Download the full version of the textbook now at textbookfull.

com

Image and Video Technology Shin'Ichi Satoh

https://textbookfull.com/product/image-and-video-
technology-shinichi-satoh/

Explore and download more textbook at https://textbookfull.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Image and Video Technology 9th Pacific Rim Symposium PSIVT

2019 Sydney NSW Australia November 18 22 2019 Proceedings
Chilwoo Lee
https://textbookfull.com/product/image-and-video-technology-9th-
pacific-rim-symposium-psivt-2019-sydney-nsw-australia-
november-18-22-2019-proceedings-chilwoo-lee/
textbookfull.com

Intelligent Image and Video Compression: Communicating

Pictures 2nd Edition David Bull

https://textbookfull.com/product/intelligent-image-and-video-
compression-communicating-pictures-2nd-edition-david-bull/

textbookfull.com

Image and video compression for multimedia engineering:

fundamentals, algorithms, and standards Third Edition Shi

https://textbookfull.com/product/image-and-video-compression-for-
multimedia-engineering-fundamentals-algorithms-and-standards-third-
edition-shi/
textbookfull.com

International finance theory and policy Eleventh Edition

Paul R. Krugman

https://textbookfull.com/product/international-finance-theory-and-
policy-eleventh-edition-paul-r-krugman/

textbookfull.com
Reviews on Biomarker Studies in Psychiatric and
Neurodegenerative Disorders Paul C. Guest

https://textbookfull.com/product/reviews-on-biomarker-studies-in-
psychiatric-and-neurodegenerative-disorders-paul-c-guest/

textbookfull.com

Normal People Sally Rooney

https://textbookfull.com/product/normal-people-sally-rooney/

textbookfull.com

Beginning Sensor Networks with XBee, Raspberry Pi, and

Arduino: Sensing the World with Python and MicroPython 2nd
Edition Charles Bell
https://textbookfull.com/product/beginning-sensor-networks-with-xbee-
raspberry-pi-and-arduino-sensing-the-world-with-python-and-
micropython-2nd-edition-charles-bell/
textbookfull.com

Children's Literature Collections: Approaches to Research

1st Edition Keith O'Sullivan

https://textbookfull.com/product/childrens-literature-collections-
approaches-to-research-1st-edition-keith-osullivan/

textbookfull.com

High-Resolution NMR Techniques in Organic Chemistry 3 ed.

3rd Edition Claridge Timothy D.W.

https://textbookfull.com/product/high-resolution-nmr-techniques-in-
organic-chemistry-3-ed-3rd-edition-claridge-timothy-d-w/

textbookfull.com
Lost with her Alien Mate A science fiction romance
Warriors of the D tali Book 6 1st Edition Ava York &
Starr Huntress [York
https://textbookfull.com/product/lost-with-her-alien-mate-a-science-
fiction-romance-warriors-of-the-d-tali-book-6-1st-edition-ava-york-
starr-huntress-york/
textbookfull.com
 Shin'ichi Satoh (Ed.)
LNCS 10799

Image and
Video Technology
PSIVT 2017 International Workshops
Wuhan, China, November 20–24, 2017
Revised Selected Papers

123
Lecture Notes in Computer Science 10799
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology Madras, Chennai, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Shin’ichi Satoh (Ed.)

Image and
Video Technology
PSIVT 2017 International Workshops
Wuhan, China, November 20–24, 2017
Revised Selected Papers

123
Editor
Shin’ichi Satoh
National Institute of Informatics
Tokyo
Japan

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-319-92752-7 ISBN 978-3-319-92753-4 (eBook)
https://doi.org/10.1007/978-3-319-92753-4

Library of Congress Control Number: 2018944410

LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics

© Springer International Publishing AG, part of Springer Nature 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

Cover illustration: Yellow Crane Pagoda, Wuhan. Photo by Reinhard Klette, Auckland, New Zealand

Printed on acid-free paper

This Springer imprint is published by the registered company Springer International Publishing AG
part of Springer Nature
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

The 8th Pacific Rim Symposium on Image and Video Technology (PSIVT 2017), held
in Wuhan, China, during November 20–24, 2017, was accompanied by a series of five
high-quality workshops covering the full range of state-of-the-art research topics in
image and video technology.
The workshops consisted of two full-day workshops and three half-day workshops
and took place on November 21. Their topics ranged from well-established areas to
novel current trends: human behavior analysis; educational cloud and image- and
video-enriched cloud services; vision meets graphics; passive and active electro-optical
sensors for aerial and space imaging; and computer vision and modern vehicles.
The workshops received 103 paper submissions (including dual submissions with
the main conference) and 36 presentations were selected by the individual workshop
committee, yielding an overall acceptance rate of 35%. The PSIVT 2017 workshop
proceedings comprise a short introduction to each workshop and all workshop con-
tributions arranged by each of the workshop organizers. We thank everyone involved
in the remarkable programs, i.e., committees, reviewers, and authors, for their distin-
guished contributions. We hope that you will enjoy reading these contributions, which
may inspire your research.

November 2017 Shin’ichi Satoh

 Contents

Human Behaviour Analysis

Deep Transfer Feature Based Convolutional Neural Forests

for Head Pose Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Yuanyuan Liu, Zhong Xie, Xi Gong, and Fang Fang

Biometric System Based on Registration of Dorsal Hand

Vein Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Szidónia Lefkovits, Simina Emerich, and László Szilágyi

A Multi-scale Triplet Deep Convolutional Neural Network

for Person Re-identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Mingfu Xiong, Jun Chen, Zhongyuan Wang, Chao Liang, Bohan Lei,
and Ruimin Hu

Facial Expression Recognition Using Cascaded Random Forest Based

on Local Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Mingjian Tuo and Jingying Chen

Detection of Salient Regions in Crowded Scenes Based on Weighted

Networks Approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Juan Zheng and Xuguang Zhang

Selecting Salient Features from Facial Components for Face Recognition. . . . 63

A. Gumede, S. Viriri, and M. V. Gwetu

Educational Cloud and Image and Video Enriched Cloud Services

Research on Optimization of Point Cloud Registration ICP Algorithm . . . . . . 81

Jian Liu, Xiuqin Shang, Shuzhan Yang, Zhen Shen, Xiwei Liu,
Gang Xiong, and Timo R. Nyberg

High School Statistical Graph Classification Using Hierarchical Model

for Intelligent Mathematics Problem Solving . . . . . . . . . . . . . . . . . . . . . . . 91
Yantao Wei, Yafei Shi, Huang Yao, Gang Zhao, and Qingtang Liu

Machine Solving on Hypergeometric Distribution Problems . . . . . . . . . . . . . 102

Chao Sun, Yao Su, and Xinguo Yu

Extracting Algebraic Relations from Circuit Images Using Topology

Breaking Down and Shrinking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Bin He, Pengpeng Jian, Meng Xia, Chao Sun, and Xinguo Yu
VIII Contents

Parallel Education Systems Under Perspective of System Construction

for New IT Era. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Xiaoyan Gong, Xiwei Liu, Sifeng Jing, and Xiao Wang

Foot Modeling Based on Machine Vision and Social

Manufacturing Research. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Hongli Peng, Zhen Shen, Xiuqin Shang, Xiwei Liu, Gang Xiong,
Taozhong Liu, and Timo R. Nyberg

Computerized Adaptive English Ability Assessment Based

on Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Xiao Wang, Yuanyuan Zhang, Shengnan Yu, Xiwei Liu,
and Fei-Yue Wang

China-Finland EduCloud Platform Towards Innovative Education . . . . . . . . . 172

Jiehan Zhou, Jukka Riekki, Mätti Hämäläinen, Pasi Mattila, Xinguo Yu,
Xiwei Liu, and Weishan Zhang

SPSE: A Smart Phone-Based Student Evaluation . . . . . . . . . . . . . . . . . . . . 186

Xiang Gao, Jiehan Zhou, Zhitao Yu, Jianli Zhao, Zhengbin Fu,
and Chunxiu Li

Research on the Construction of Corpus for Automatic Solution

of Elementary Mathematics Statistics Applications . . . . . . . . . . . . . . . . . . . 194
Chuanyuan Lao, Qingtang Liu, Linjing Wu, Jingxiu Huang,
and Gang Zhao

Constructing a Learning Map with Lattice for a Programming Course . . . . . . 204

Xin Li, Han Lyu, Jiehan Zhou, Shuai Cao, and Xin Liu

Data Mining as a Cloud Service for Learning Artificial Intelligence . . . . . . . 214

Weishan Zhang, Hao Lv, Liang Xu, Xin Liu, and Jiehan Zhou

Vision Meets Graphics

Crowd Counting from a Still Image Using Multi-scale Fully Convolutional

Network with Adaptive Human-Shaped Kernel . . . . . . . . . . . . . . . . . . . . . . 227
Jinmeng Cao, Biao Yang, Yuyu Zhang, and Ling Zou

On Road Vehicle Detection Using an Improved Faster RCNN Framework

with Small-Size Region Up-Scaling Strategy . . . . . . . . . . . . . . . . . . . . . . . 241
Biao Yang, Yuyu Zhang, Jinmeng Cao, and Ling Zou

Fast Haze Removal of UAV Images Based on Dark Channel Prior . . . . . . . . 254
Siyu Zhang, Congli Li, and Song Xue

Watercolour Rendering of Portraits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

Paul L. Rosin and Yu-Kun Lai
 Contents IX

Blind Image Deblurring via Salient Structure Detection and Sparse

Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Yu Cai, Jinshan Pan, and Zhixun Su

Blur Estimation for Natural Edge Appearance in Computational

Photography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
Dongwei Liu and Reinhard Klette

Structure-Preserving Texture Smoothing via Adaptive Patches . . . . . . . . . . . 311

Hui Wang, Yue Wang, Junjie Cao, and Xiuping Liu

Robust Blind Deconvolution Using Relative Total Variation as a

Regularization Penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
Yunzhi Lin and Wenze Shao

Passive and Active Electro-Optical Sensors for Areal

and Space Imaging

An Aircraft Tracking Method in Simulated Infrared Image Sequences

Based on Regional Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Sijie Wu, Saisai Niu, Kai Zhang, and Jie Yan

DESIS - DLR Earth Sensing Imaging Spectrometer. . . . . . . . . . . . . . . . . . . 356

David Krutz, Holger Venus, Andreas Eckardt, Ingo Walter,
Ilse Sebastian, Ralf Reulke, Burghardt Günther, Bernd Zender,
Simone Arloth, Christian Williges, Matthias Lieder, Michael Neidhardt,
Ute Grote, Friedrich Schrandt, and Andreas Wojtkowiak

FireBIRD Mission Data for Gas Flaring Analysis . . . . . . . . . . . . . . . . . . . . 369

Agnieszka Soszyńska

Automatic Ship Detection on Multispectral and Thermal Infrared

Aerial Images Using MACS-Mar Remote Sensing Platform . . . . . . . . . . . . . 382
Jörg Brauchle, Steven Bayer, and Ralf Berger

Extracting Plücker Line and Their Relations for 3D Reconstruction

of Indoor Scene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
Huihui Sun, Xinguo Yu, and Chao Sun

Computer Vision and Modern Vehicles

Context-Awareness Based Adaptive Gaussian Mixture

Background Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
HongGang Xie, JinSheng Xiao, and JunFeng Lei

Robust Expression Recognition Using ResNet with a Biologically-Plausible

Activation Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
Yunhua Chen, Jin Du, Qian Liu, and Bi Zeng
X Contents

Local Fast R-CNN Flow for Object-Centric Event Recognition

in Complex Traffic Scenes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Qin Gu, Jianyu Yang, Wei Qi Yan, Yanqiang Li, and Reinhard Klette

Mixed-Noise Removal in Images Based on a Convolutional

Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
Ling Ding, Huyin Zhang, Bijun Li, Jian Zhou, and Wenhao Gu

A Systematic Scheme for Automatic Airplane Detection

from High-Resolution Remote Sensing Images . . . . . . . . . . . . . . . . . . . . . . 465
Jiao Zhao, Jing Han, Chen Feng, and Jian Yao

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479

Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
Human Behaviour Analysis
 Workshop on Human Behavior Analysis

With the rapid development of image and video technologies, it became possible to
analyze human behavior via nonintrusive sensors. This endows computers with a
capacity to understand what people are doing, the things they are interested in, their
preference and personality in a nonintrusive manner. Human behavior analysis has long
been a critical issue in developing human-centered multimedia interaction systems
including affect-sensitive systems with educational goals, intelligent surveillance,
gesture interaction system, and smart tutoring, etc.
This workshop received 18 submissions about using image and video technologies
for human behavior analysis, including action and activity recognition, affect and
attention analysis, social signal processing, face analysis, gestures interaction, intelli-
gent surveillance, and smart tutoring. After a rigorous peer reviewing, this workshop
accepted six submissions. The workshop was held on November 21, 2017, and all the
accepted papers were presented.
 Organization

Workshop Committee

Jingying Chen (Co-chair) China Central Normal University, China

Dan Chen (Co-chair) Wuhan University, China
Bernard Tiddeman Aberystwyth University, UK
Suya You University of Southern California, USA
Liyuan Li Institute for Infocomm Research, Singapore
Joanna Kolodziej University of Bielsko-Biała, Poland
Rajiv Ranjan Newcastle University, UK
Xun Xu National University of Singapore, Singapore
Jiayi Ma Wuhan University, China
Changxin Gao Huazhong University of Science and Technology, China
Quan Zhou Nanjing University of Posts and Telecommunications,
China
Jingang Yu South China University of Technology, China
Bin Sun University of Electronic Science and Technology
of China, China
Leyuan Liu China Central Normal University, China
Kun Zhang China Central Normal University, China
 Deep Transfer Feature
Based Convolutional Neural Forests
for Head Pose Estimation

Yuanyuan Liu, Zhong Xie, Xi Gong(B) , and Fang Fang

Faculty of Information Engineering, China University of Geosciences,

Wuhan 430074, China
gongxi [email protected]

Abstract. In real-world applications, factors such as illumination,

occlusion, and poor image quality, etc. make robust head pose estima-
tion much more challenging. In this paper, a novel deep transfer feature
based on convolutional neural forest method (D-CNF) is proposed for
head pose estimation. Deep transfer features are extracted from facial
patches by a transfer network model, firstly. Then, a D-CNF is devised
to integrate random trees with the representation learning from deep con-
volutional neural networks for robust head pose estimation. In the learn-
ing process, we introduce a neurally connected split function (NCSF)
as the node splitting strategy in a convolutional neural tree. Experi-
ments were conducted using public Pointing’04, BU3D-HP and CCNU-
HP facial datasets. Compared to the state-of-the-art methods, the pro-
posed method achieved much improved performance and great robust-
ness with an average accuracy of 98.99% on BU3D-HP dataset, 95.7%
on Pointing’04 and 82.46% on CCNU-HP dataset. In addition, in con-
trast to deep neural networks which require large-scale training data,
our method performs well even when there are only a small amount of
training data.

Keywords: Convolutional neural network · Random forest

Transfer network · Head pose estimation

1 Introduction
Head pose estimation is the key step in many computer vision applications,
such as human computer interaction, intelligent robotics, face recognition, and
recognition of visual focus of attention [14,28]. The existing techniques achieve
satisfactory results in well-designed environments. In real-world applications,
however, factors, such as illumination variation, occlusion, poor image quality,
etc., make head pose estimation much more challenging [19,22]. Hence, we pro-
pose a deep transfer feature based convolutional neural forest (D-CNF) method
to estimate head pose estimation in unconstrained environment.

c Springer International Publishing AG, part of Springer Nature 2018

S. Satoh (Ed.): PSIVT 2017, LNCS 10799, pp. 5–16, 2018.
https://doi.org/10.1007/978-3-319-92753-4_1
6 Y. Liu et al.

A general head pose estimation framework appeared in most of previous

works can be divided into two major steps, one is the feature extraction and
the other is classifier construction [13]. Extracting robust facial features and
designing effective classifier are the two key factors in unconstrained head pose
estimation. For feature extraction, based on different features, several methods
for the problem can be briefly divided into two categories, facial local feature and
facial global feature based methods. The former methods usually require high
image resolution for facial local feature identification, such as eyes, eyebrows,
nose or lips [13,27], etc. These methods can provide accurate recognition results
relying on accurate detection of facial feature points and high quality images. The
latter methods based on facial global feature usually use texture features from
an entire face to estimate head poses [1,4,17], etc. It may be good for dealing
with low resolution image but not robust to occlusion and illumination. In the
real-life scene, the various illumination occlusion, low image resolution and wide
scene make facial local feature extraction difficult. In order to extract robust
high-level features for head pose estimation, we address the problem based on
globe deep transfer feature representation.
For the head pose classifier construction, most of the traditional classifiers,
such as Support vector machine (SVM), Random forest (RF), Bays classifier and
convolutional neural network (CNN), together with some unsupervised learning
techniques are employed in the head pose estimation [17,21,25]. Recent years,
CNN and RF become popular learning algorithms for head pose estimation in
some real-life applications. CNN has an ability to automatically learn high-level
feature representations from raw image data [11,16,20,24,30]. CNN achieves
huge success in face recognition [23] and object multi-classification [26]. However,
a limit for CNN is that the learning procedure needs a large amount of datasets
and GPUs [6,9,15]. RF is a popular method given their capability to handle
large training datasets, high generalization power and speed, and easy imple-
mentation [2,3,5,7]. In this paper, we are interested in constructing an effective
head pose classifier using a limited amount of image data with a hybrid deep
convolution networks enhanced decision forest. Our method aims at improving
both accuracy and efficiency. The pipeline of our proposed D-CNF is depicted in
Fig. 1. The deep transfer feature is extracted by transfer CNN model to suppress
the influence of illumination, occlusion, and low image resolution, firstly. Then,
head poses are estimated by the trained D-CNF model.

Fig. 1. The pipeline of D-CNF for head pose estimation

 Deep Transfer Feature Based Convolutional Neural Forests 7

Our contributions include the following:

1. We propose a deep transfer feature based convolutional neural forest method

(D-CNF) for head pose estimation in unconstrained environment, which uni-
fies classification trees with the representation learning from deep convolution
networks, by training them in an end-to-end way.
2. We introduce a neurally connected split function (NCSF) as new split node
learning in a D-CNF tree. The D-CNF method can achieve fast and accurate
recognized results in the limited amount of image data, rather than a large
amount of data by CNN.
3. We propose a robust deep transfer feature representation based on a pre-
trained CNN model.

The rest of this paper is organized as follows: Sect. 2 presents our D-CNF
method in details. Section 3 discusses the experimental results using publicly
available datasets. Section 4 concludes this paper with a summary of our method.

2 Deep Transfer Feature Based Convolution Neural

Forests for Head Pose Estimation

In this section, we address the D-CNF approach for head pose estimation in
unconstrained environment. First, we present robust deep feature representation
based on facial patches, which can reduce the influence of various noises, such
as over-fitting, illumination, low image resolution, etc. Then, we describe the
framework of D-CNF training procedure for head pose estimation in details.
Finally, we give the D-CNF prediction for head pose estimation in unconstrained
environment.

2.1 Deep Transfer Feature Representation

We extract deep transfer feature from facial patches with a pre-trained CNN
model, i.e., Vgg-face [23]. We employ the Vgg-face architecture that is pre-
trained with the LFW and YTF face datasets [23] to derive deep high-level
feature representation, as shown in Fig. 2. The model includes 13 convolution
layers, 5 max-pooling layers, and 3 fully connected layers. The deep transfer
feature is described as:

y j = max(0, xi wi,j + bj ), (1)
i

where y j is the j th output feature value of the first fully connected layer, xi
is the ith feature map of the last convolution layer, wi,j indicates the weight
between the ith feature map and the j th output feature value, and bj donates
the bias of the j th output feature value. The deep transfer feature is used to train
a two-layer network through back propagation, which can transfer the original
Vgg-face feature to the pose feature.
8 Y. Liu et al.

Fig. 2. The structure of pre-trained CNN network for deep feature representation. The
trained network model includes 13 convolution layers, 5 max-pooling layers, and 3 full
connection layers. In our work, we extract deep features from facial patches on the first
connection layer.

2.2 D-CNF Training

In this paper, we propose a fast and efficient D-CNF method for robust head
pose estimation on limit training sets, which is unifies classification trees with
the representation learning from deep convolution networks, by training them
in end-to-end way. The training of a traditional decision tree of a random forest
(RF) consists in a recursive procedure, which starts from the root and iteratively
builds the tree by splitting nodes [2]. The proposed D-CNF is also an ensemble of
convolution neural trees, where split nodes are computed by the proposed neural
connected split function (NCSF). The proposed NCSF can improve the learn-
ing capability of splitting node by deep neural learning representation, thus to
improve the discrimination and efficiency of a tree. The detail training procedure
is given as below.

Learning Splitting Nodes by NCSF. For facial patches, we extract a set

of deep transfer features P , P = {Pi } and Pi = {y j }. We propose a NCSF-fn
to reinforce the learning capability of a splitting node by deep neural learning
representation. Each output of fn is brought in correspondence with a splitting
node dn (Pi ; Y ),
dn (P ; Y ) = σ(fn (P ; Y )), (2)
where σ(x) = (1 + e−x )−1 is the sigmoid function and Y is the decision node
parametrization.
We employ a Stochastic Gradient Descent (SGD) approach to minimize the
risk with respect to Y :
 ∂L (t)
Y (t+1) = Y (t) − |B|
η
∂Y (Y , π; P ), (3)
(P,π)∈B

where η > 0 is the learning rate, π is facial expression label and B is a random
subset (a.k.a. mini-batch) of samples. The gradient with respect to Y is obtained
by chain rule as follows:

∂L(Y, π; P )  ∂L(Y, π; P ) ∂fn (P ; Y )

= · . (4)
∂Y ∂fn (P ; Y ) ∂Y
n∈N
 Deep Transfer Feature Based Convolutional Neural Forests 9

Hence, the gradient term that depends on the neutral decision tree is
∂L(Y, π; P )
= −(dNr Nl
n (P ; Y ) + dn (P ; Y )), (5)
∂fn (P ; Y )
where given a node N in a tree and Nr and Nl denote its right and left child,
respectively.
To split a node, Information Gain (IG) is maximized:
 S
 d n 
ϕ̃ = arg max(H(dn ) − (H(dn ))), (6)
ϕ S∈{Nr ,Nl } |dn |

|dS |
where |dnn | , s ∈ {Nr , Nl } is the ratio between the number of samples in dNl
n (arriv-
ing at the left child node), set dN r
n (arriving at the right child node), and H(dn )
is the entropy of dn .

Learning Leaf Nodes. Create a leaf l when IG is below a predefined threshold

or when a maximum depth is reached. Otherwise continue recursively for the two
child nodes dNl Nr
n and dn at the splitting node step. For a leaf node in a conditional
D-CNF tree, it stores the conditional multi-probability p(π|θ, y). Therefore, we
simplify the distribution over head poses by a multivariate Gaussian Mixture
Model (GMM) [17] as in:

p(θ, l) = N (θ; θ, Σlθ ), (7)

where θ and Σlθ are the mean and covariance of leaves’ head pose probabilities,
respectively.

2.3 D-CNF for Head Pose Estimation

This section provides the prediction procedure of the D-CNF for head pose
estimation. Deep transfer feature patches pass through the trees in a trained
D-CNF. All feature patches end in a set of leaves of the forest. In the leaves of
a D-CNF forest, there are muti-probabilistic models of head poses. We simplify
the distributions over multi-probabilities by adopting multivariate GMM as:

p(θ|l) = N (θ; θ, Σlθ ), (8)

where θ and Σlθ are the mean and covariance of leaves’ head pose probabilities,
respectively.
While Eq. 8 models the probability for a feature patch pi ending in the leaf
l of a single tree, the probability of the forest is obtained by averaging over all
trees:
1
p(θ|P ) = p(θ|lt (P )) (9)
T t
where lt is the corresponding leaf for the tree Tt , T is the number of trees in
D-CNF.
10 Y. Liu et al.

3 Experimental Results
3.1 Datasets and Settings
To evaluate our approach, three challenging face datasets were used: Pointing’04
dataset [10], BU3D-HP dataset [31], and CCNU-HP dataset in the wide class-
room [17]. These datasets were chosen since they contained unconstrained face
images with poses ranging from −90◦ to +90◦ . The Pointing’04 head pose dataset
is a benchmark of 2790 monocular face images of 15 people with variations of
yaw and pitch angles from −90◦ to +90◦ . For every person, 2 series of 93 images
(93 different poses) are available. The CCNU dataset was collected included an
annotated set of 38 people with 75 different head poses from an overhead camera
in the wide scene. It contains head poses spanning from −90◦ to 90◦ in horizon-
tal direction, and −45◦ to 90◦ in vertical direction. The multi-view BU3D-HP
database contains 100 people of different ethnicities, including 56 females and
44 males with variations of yaw angles from −90◦ to +90◦ .

Fig. 3. The examples of head pose estimation on Pointing’04, BU3D-HP and CCNU-
HP datasets. Top row: results of Pointing’04. Middle row: results of the BU3D-HP
dataset. Bottom row: results of CCNU-HP dataset.

The examples of head pose estimation on Pointing’04, BU3D-HP and CCNU-

HP datasets are shown in Fig. 3. The D-CNF method can achieve fast and accu-
rate recognized results in limited amount of image data, rather than a large
amount of data by CNN. Our method was trained with 2000 images from Point-
ing’04 dataset, 15498 images from BU3D-HP dataset and 2121 images from
CCNU-HP dataset. In evaluation, we used 870 images from Pointing’04 dataset,
5166 images from BU3D-HP dataset and 707 images from CCNU-HP dataset.
The experiments were conducted in a PC with Intel(R) Core(TM) i7-6700 CPU@
 Deep Transfer Feature Based Convolutional Neural Forests 11

4.00 GHz, RAM 32 GB, NVIDIA GeForce GTX 1080 (2). We use the Caffe frame-
work [12] for the transfer CNN and deep feature representation.

3.2 Experiments on Pointing’04 Datasets

Figure 4 shows the head poses estimation results on Pointing’04 datasets in yaw
and pitch rotations, respectively. The average accuracy on 9 yaw head poses and
9 pitch head poses is 95.6%. As it is shown, the highest accuracy is 98.4% of 90◦
in the yaw rotation. The lowest accuracy is 92.6% of −45◦ in the pitch rotation,
due to more occlusion in a face area.

Fig. 4. Head pose estimation on Pointing’04 datasets in the yaw and pitch rotations

In comparison with the state-of-the-art head pose estimation methods, we

conducted experiments using the MSHF [18], Multivariates label distribution
(MLD-wj) [29], CNN(6convs+2fc) [15], multi-class SVM (M-SVM) [22] and
HF [8] on Pointing’04 head pose dataset. The same training and testing datasets
were used, and we employed a 4-fold cross-validation. Table 1 lists the average
accuracy and error across using these methods. MLD-wj [29], CNN [15] and HF
[8] yielded comparable results with an accuracy of approximately 70% in yaw
and pitch rotations. MLD-wj [29] proposed to associate a multivariate label dis-
tribution to each image for head pose estimation in yaw and pitch rotations.
MSHF [18] proposed a hybrid structure hough forest to 25 class head pose esti-
mation and achieved the second highest accuracy of 84%. HF [8] improved ran-
dom forests with Hough voting for real-time head pose estimation. M-SVM [22]
produced similar accuracy in the range of 60%. Our proposed D-CNF exhibited
the best performance with the accuracy of 95.7% in yaw and pitch rotations.
In addition, the standard deviation of D-CNF indicates that D-CNF achieved
the greatest consistency with a smallest STD. It is evidential that our D-CNF
improved the head pose estimation with great robustness.
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
12 Y. Liu et al.

Table 1. Accuracy (%) and average error (in degrees) using different methods on
Pointing’04 dataset.

Methods Yaw Pitch Yaw + Pitch STD

MSHF [18] 92.3 90.7 84.0 3.5
MLD-wj [29] 84.30 86.24 72.3 4.9
CNN [15] 83.52 86.94 71.83 5.5
HF [8] 82.3 84.86 70.54 5.2
SVM [22] 80.6 82.5 60.46 5.7
D-CNF 99.05 94.36 95.7 0.8

3.3 Experiments on Multi-view BU3D-HP Dataset

Each image in the BU3D-HP dataset is automatically annotated with one out
of the nine head pose labels ({−90◦ , −60◦ , −45◦ , −30◦ , 0◦ , +30◦ , +60◦ , +75◦ ,
90◦ }). We train a D-CNF of 50 neural trees using 15498 head pose images.
Figure 5 shows the confusion matrix of head pose estimation on BU3D-HP
dataset. The D-CNF estimated 9 head pose classes in the horizontal direction
and achieved the average accuracy of 98.99%. Examples of the estimated head
pose are shown in Fig. 3.

Fig. 5. Confusion matrix of head pose estimation on BU3D-HP dataset.

The average accuracy of our D-CNF method is compared with that of CNN,
Zheng GSRRR [33], and SIFT + CNN [32] in Table 2. The CNN in this experi-
ment contains three convolution layers followed by three max-pooling layers and
two fully connected layers. Each filter is of size 5×5 and there are 32, 64, and 128
such filters in the first three layers, respectively. The input images are rescaled
to 224 by 224.
The accuracy of the CNN on BU3D-HP dataset is 69.61% as presented in
Table 2. The accuracies achieved with SIFT using algorithms proposed in [32,
33] are 87.36% and 92.26%, respectively. Our method achieves 98.99% which is
 Deep Transfer Feature Based Convolutional Neural Forests 13

Table 2. Accuracy (%) and STD using different methods on multi-view BU3D-HP
dataset.

Methods Features Poses Accuracy STD.

CNN Image 9 69.61 0.9
Zheng GSRRR [33] Sparse SIFT 9 87.36 0.8
SIFT + CNN [32] SIFT 9 92.26 0.7
D-CNF Deep transfer feature 9 98.99 0.5

competitive to the methods above. The lowest STD. of 0.5% using our method
also proved the robustness of the proposed D-CNF.

3.4 Experiments on CCNU-HP Dataset in the Wide Scene

In this case, we evaluated the proposed D-CNF on CCNU-HP dataset in the
wide scene. For evaluation, a 4-fold cross-validation was conducted. In our exper-
iments, we annotate the dataset into 5 classes in the yaw rotation as Fig. 6(a)
and 4 classes in the pitch rotation as Fig. 6(b). The final classified classes are 20
categories in the wide scene dataset.

Fig. 6. The annotation categories of the yaw and pitch angels in the experiments. (a)
The annotation classes in the yaw rotation, (b) The annotation classes in the yaw
rotation.

Figure 7 shows the confusion matrixs of head pose estimation on CCNU-

HP dataset in the yaw and pitch rotations, respectively. The D-CNF achieved
the average accuracy of 88.54% in the yaw rotation and 76.38% in the more
challenging pitch rotation. Examples of the estimated head pose are shown in
Fig. 3.
Table 3 lists the average accuracy and error across on more challenging
CCNU-HP datasets using four state-of-the-art methods. The average accuracy
of the CNN on CCNU-HP dataset is 59.52% as presented in Table 3. The second
highest accuracy is achieved 77.9% with combined features using D-RF method.
Our method achieves 82.46% which is competitive to the methods above.
14 Y. Liu et al.

Fig. 7. Confusion matrixs of head pose estimation on CCNU-HP dataset. (a) The
matrix of yaw angles, (b) The matrix of pitch angles.

Table 3. Accuracy (%) using different methods on CCNU-HP dataset.

Methods Features Yaw Pitch Yaw + Pitch

CNN Image 65.25 53.79 59.52
Gabor + RF Gabor 75.42 67.57 71.5
D-RF [17] Combined features 85.6 70.19 77.90
D-CNF Deep transfer feature 88.54 76.38 82.46

4 Conclusion

This paper described a novel deep transfer feature based convolutional neural
enhanced forests (D-CNF) method for head pose estimation in unconstrained
environment. In this method, robust deep transfer features are extracted from
facial patches using transfer CNN model, firstly. Then, the D-CNF integrates
random trees with the representation learning from deep convolutional neural
networks for head pose estimation. Besides, a neural connected split function
(NCSF) is introduced to D-CNF to split node learning. Finally, a prediction
procedure of the trained D-CNF can classify head pose in unconstrained envi-
ronment. Our method can perform well in limit number of datasets owing to
transferring pre-trained CNN to fast decision node splitting in a Random For-
est. The experiments demonstrate that our method has remarkable robustness
and efficiency.
Experiments were conducted using public Pointing’04, BU3D-HP and
CCNU-HP datasets. Our results demonstrated that the proposed deep feature
outperformed the other popular image features. Compared to the state-of-the-
art methods, the proposed D-CNF achieved improved performance and great
robustness with an average accuracy of 98.99% on BU3D-HP dataset, 95.7% on
Pointing’04 dataset, and 82.46% on CCNU-HP dataset. The average time for
performing a head pose estimation is about 113 ms.
Compared to CNN method from popular deep learning, our method achieved
the greatest performance on limited number of datasets. In future, we plan to
investigate on-line learning methods to achieve real-time estimation by integrat-
ing head movement tracking.
 Deep Transfer Feature Based Convolutional Neural Forests 15

Acknowledgments. This work was supported by the National Natural Science

Foundation of China (No. 61602429), China Postdoctoral Science Foundation (No.
2016M592406), and Research Funds of CUG from the Colleges Basic Research and
Operation of MOE (No. 26420160055).

References
1. Ahn, B., Park, J., Kweon, I.S.: Real-time head orientation from a monocular cam-
era using deep neural network. In: Cremers, D., Reid, I., Saito, H., Yang, M.-H.
(eds.) ACCV 2014. LNCS, vol. 9005, pp. 82–96. Springer, Cham (2015). https://
doi.org/10.1007/978-3-319-16811-1 6
2. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
3. Bulo, S.R., Kontschieder, P.: Neural decision forests for semantic image labeling.
In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 81–88
(2014)
4. Chu, X., Ouyang, W., Li, H., Wang, X.: Structured feature learning for pose esti-
mation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 4715–4723 (2016)
5. Dantone, M., Gall, J., Fanelli, G., Van Gool, L.: Real-time facial feature detection
using conditional regression forests. In: 2012 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pp. 2578–2585. IEEE (2012)
6. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.:
Decaf: a deep convolutional activation feature for generic visual recognition. In:
ICML, vol. 32, 647–655 (2014)
7. Fanelli, G., Yao, A., Noel, P.-L., Gall, J., Van Gool, L.: Hough forest-based facial
expression recognition from video sequences. In: Kutulakos, K.N. (ed.) ECCV 2010.
LNCS, vol. 6553, pp. 195–206. Springer, Heidelberg (2012). https://doi.org/10.
1007/978-3-642-35749-7 15
8. Garcı́a-Montero, M., Redondo-Cabrera, C., López-Sastre, R., Tuytelaars, T.: Fast
head pose estimation for human-computer interaction. In: Paredes, R., Cardoso,
J.S., Pardo, X.M. (eds.) IbPRIA 2015. LNCS, vol. 9117, pp. 101–110. Springer,
Cham (2015). https://doi.org/10.1007/978-3-319-19390-8 12
9. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE International Conference on
Computer Vision, pp. 1440–1448 (2015)
10. Gourier, N., Hall, D., Crowley, J.: Estimating face orientation from robust detec-
tion of salient facial features in pointing. In: International Conference on Pattern
Recognition Workshop on Visual Observation of Deictic Gestures, pp. 1379–1382
(2004)
11. Insafutdinov, E., Pishchulin, L., Andres, B., Andriluka, M., Schiele, B.: DeeperCut:
a deeper, stronger, and faster multi-person pose estimation model. In: Leibe, B.,
Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9910, pp. 34–50.
Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46466-4 3
12. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadar-
rama, S., Darrell, T.: Proceedings of the 22nd ACM International Conference on
Multimedia
13. Wu, J., Trivedi, M.M.: A two-stage head pose estimation framework and evaluation.
Pattern Recogn. 41, 1138–1158 (2008)
14. Kim, H., Sohn, M., Kim, D., Lee, S.: Kernel locality-constrained sparse coding for
head pose estimation. IET Comput. Vis. 10(8), 828–835 (2016)
16 Y. Liu et al.

15. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep con-
volutional neural networks. In: Advances in Neural Information Processing Sys-
tems, pp. 1097–1105 (2012)
16. Liu, X., Liang, W., Wang, Y., Li, S., Pei, M.: 3D head pose estimation with convo-
lutional neural network trained on synthetic images. In: 2016 IEEE International
Conference on Image Processing (ICIP), pp. 1289–1293. IEEE (2016)
17. Liu, Y., Chen, J., Shu, Z., Luo, Z., Liu, L., Zhang, K.: Robust head pose estimation
using dirichlet-tree distribution enhanced random forests. Neurocomputing 173,
42–53 (2016)
18. Liu, Y., Xie, Z., Yuan, X., Chen, J., Song, W.: Multi-level structured hybrid forest
for joint head detection and pose estimation. Neurocomputing 266, 206–215 (2017)
19. Ma, B., Li, A., Chai, X., Shan, S.: CovGa: a novel descriptor based on symmetry
of regions for head pose estimation. Neurocomputing 143, 97–108 (2014)
20. Mukherjee, S.S., Robertson, N.M.: Deep head pose: gaze-direction estimation in
multimodal video. IEEE Trans. Multimedia 17(11), 2094–2107 (2015)
21. Murphy-Chutorian, E., Trivedi, M.M.: Head pose estimation in computer vision:
a survey. IEEE Trans. Pattern Anal. Mach. Intell. 31(4), 607–626 (2009)
22. Orozco, J., Gong, S., Xiang, T.: Head pose classification in crowded scenes. In:
British Machine Vision Conference, London, UK, pp. 1–3, 7–10 September 2009
23. Parkhi, O.M., Vedaldi, A., Zisserman, A.: Deep face recognition. In: BMVC, vol.
1, p. 6 (2015)
24. Patacchiola, M., Cangelosi, A.: Head pose estimation in the wild using convolu-
tional neural networks and adaptive gradient methods. Pattern Recogn. 71, 132–
143 (2017)
25. Ranjan, R., Patel, V.M., Chellappa, R.: Hyperface: a deep multi-task learning
framework for face detection, landmark localization, pose estimation, and gender
recognition. arXiv preprint arXiv:1603.01249 (2016)
26. Rastegari, M., Ordonez, V., Redmon, J., Farhadi, A.: XNOR-Net: imagenet classi-
fication using binary convolutional neural networks. In: Leibe, B., Matas, J., Sebe,
N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9908, pp. 525–542. Springer, Cham
(2016). https://doi.org/10.1007/978-3-319-46493-0 32
27. Schwarz, A., Lin, Z., Stiefelhagen, R.: HeHOP: highly efficient head orientation
and position estimation. In: 2016 IEEE Winter Conference on Applications of
Computer Vision (WACV), pp. 1–8. IEEE (2016)
28. Wu, S., Kan, M., He, Z., Shan, S., Chen, X.: Funnel-structured cascade for
multi-view face detection with alignment-awareness. Neurocomputing 221, 138–
145 (2017)
29. Xin, G., Xia, Y.: Head pose estimation based on multivariate label distribution.
In: IEEE Conference on Computer Vision and Pattern Recognition, Ohio, USA,
pp. 1837–1842, 24–27 June 2014
30. Xu, X., Kakadiaris, I.A.: Joint head pose estimation and face alignment framework
using global and local CNN features. In: Proceedings of the 12th IEEE Conference
on Automatic Face and Gesture Recognition, Washington, DC, vol. 2 (2017)
31. Yin, L., Wei, X., Sun, Y., Wang, J., Rosato, M.J.: A 3D facial expression database
for facial behavior research. In: 2006 7th International Conference on Automatic
Face and Gesture Recognition, FGR 2006, pp. 211–216. IEEE (2006)
32. Zhang, T., Zheng, W., Cui, Z., Zong, Y., Yan, J., Yan, K.: A deep neural network-
driven feature learning method for multi-view facial expression recognition. IEEE
Trans. Multimedia 18(12), 2528–2536 (2016)
33. Zheng, W.: Multi-view facial expression recognition based on group sparse reduced-
rank regression. IEEE Trans. Affect. Comput. 5(1), 71–85 (2014)
Random documents with unrelated
content Scribd suggests to you:
happen that the original photograph would re-develop on top of the
second, very careful chemical cleaning of the plate being necessary
to prevent this. Mumler’s first spirit photograph was probably
produced in this way, and the knowledge was turned to good
account by several of the earlier spirit photographers. Some of the
unexpected results obtained by amateurs may be attributable to this
cause, because a certain number of used plates are returned to
plate manufacturers, who clean off the emulsion and use the glass
again. The cleansing may sometimes be imperfect, and in these
cases the original image may appear on development.
2. Faces may be sketched in chemicals on small pieces of card, or
even on the medium’s fingers. On opportunity arising in the dark-
room, the medium holds or steadies the plate for an instant,
bringing the chemical pictures into contact with the plate. Or he may
so manœuvre it that the plate is laid face down on a prepared
surface of the dark-room work-bench, probably while it is being
marked[7]; upon development of the plate extras will duly appear.
The most refined version of this method consists in the preparation
of small rubber stamps in which the chemicals are smeared. These
can easily be palmed and dabbed for a moment on the plate in a
manner which appears quite unsuspicious. A number of active
chemicals will produce this effect, but the medium must be careful
to know whether the substance he is using will accelerate or retard
development in the affected part; for cases have occurred in which a
positive extra has been produced on the negative plate, giving a
negative spirit on the finished print!
3. Mr. Bush, in his recent pamphlet, “Spirit Photography Exposed,”
describes a piece of apparatus made out of an empty blacking-tin
containing a small electric bulb, one side of the tin being replaced by
a positive transparency of the desired extra. This, he alleges, is used
by Hope, the Crewe spirit photographer, the transparency being
pressed against the plate and the light switched on for a second. If
carefully faced with black velvet round the transparency, this device
should be quite useful; but it must be remembered that an escaping
ray of white light would at once catch the eye in the dark-room.
Skilful palming and manipulation should make it quite possible for an
extra to be printed on the plate in this way, if the medium can cover
it with his hand for a moment or two. All Hope’s results are certainly
not produced in this way, however, as is implied by Mr. Bush.
4. The medium may palm a positive transparency; if he is allowed to
handle the plate he will hold it close to the red lamp with the
transparency between; if the lamp is rather bright, or is not a very
deep red, an impression is soon made on the plate.
5. With a pinhole in the dark-room lamp, and a transparency inside—
a perfectly practicable arrangement with some of the more
complicated dark-room “safe-lights,”—a pinhole projector can be
formed, which will throw an image on a suitably-placed plate. Any
leakage of white light into the dark-room, either from the lamp or
from outside, can be used to produce blotches and streaks on the
plate. A very little mechanical ingenuity will enable a medium who
takes a pride in his work to rig up an arrangement of this kind which
can be switched off and on at will and which will project an image
on a predetermined spot on the bench. By the simple expedient of
having the bench so cluttered up with bottles and miscellaneous
rubbish that this spot is the only unencumbered one, the
unsuspecting sitter may be forced to lay a plate on this spot while,
for example, he is marking another. The medium may ostentatiously
stand at the other end of the room and “switch on” for a moment
while the sitter’s attention is engaged with his marking.
6. Photographic plates are sensitive to rays invisible to the eye, as
has been pointed out in considering the effect of fluorescent
substances. X-rays and ultra-violet rays, for instance, both invisible
yet strongly actinic, might be used in the most baffling manner in
the production of spirit extras. The expense and technical difficulties
would be considerable, but were any medium to take the method
up, he might safely defy the most critical investigation and would
soon recoup himself for the few pounds initial outlay.
There are undoubtedly many other methods used by mediums for
this purpose; but if the sitter who has obtained spirit extras under
test conditions carefully considers the procedure employed, in the
light of the suggestions made above, he will probably find that
several loopholes were left by which fraud might have been
introduced.

B.—Experiments in Fraud
The argument most frequently brought forward, in favour of the
genuineness of spirit photographs, is that the conditions employed in
their taking leave no loophole for fraud. It has been pointed out in
the preceding section that the usual “test conditions” leave not one,
but many, such loopholes. Evidence of fraud has at some time or
other been brought against most spirit photograph mediums, and
they have consequently been more or less discredited. Other
mediums have been more clever—or more fortunate—and many
people therefore argue that they are not all to be tarred with the
same brush; it is pointed out that spirit extras have been obtained
under the strictest conditions imposed by acute observers who have
found nothing suspicious of trickery.
It occurred to me that the most effective way to refute this
argument was actually to produce bogus spirit photographs under
similar, or even more stringent, test conditions. This I accordingly
attempted in a series of séances, held in my rooms at Cambridge in
the summer of 1919. At four of these séances photographs were
taken, and on each occasion one plate showed a more or less
conventional spirit extra. As I was experimenting primarily for my
own satisfaction, my seven victims were drawn from among my own
friends, and were enjoined to keep the matter as quiet as possible.
They were not, of course, specially trained psychic researchers, but
could not, I think, be considered as being particularly easy men to
deceive. Five of the seven were ex-Service men, and all were of B.A.
or “fourth year” University status; they included two chemists, two
medical students, a geologist, and two physiologists who were also
studying psychology. They were all therefore of a scientific bent,
and, with possibly one exception, were completely sceptical about
spiritualistic phenomena when the experiments started.
I first suggested to four of them that we might try to obtain a spirit
photograph, like those described and reproduced in recent magazine
articles. They did not take me very seriously at first, but after we
had obtained the right atmosphere with a little table-turning, they
consented to try for a spirit photograph. When a spirit face duly
developed in addition to the sitter, everyone present expressed
amazement! I was naturally asked if I was “pulling their legs.” I
hedged and refused to say either yes or no, explaining that I wanted
the experiments to continue under scientific conditions. If, on the
one hand, I declared that I had not in any way faked the
photograph, they would probably believe me, and would not insist
on further photographs being taken under test conditions. If, on the
other hand, I refused to give such an assurance, they would think
that I was probably tricking them, and would take all possible steps
to “bowl me out”; and when they failed to do so would thereby
establish evidence of the genuineness of any further photographs we
might be lucky enough to obtain. After some little demur they saw
the point of this—or as much of it as I wished them to see—and
agreed to meet again in my room on the following Sunday evening,
promising that I should be given no opportunity of playing any
tricks. It was also agreed that notes should be taken during the
séances as far as possible, and that full reports of what occurred
should be drawn up afterwards by all of us in conjunction, which
everyone would sign.
I now quote their report on the next two meetings, omitting nothing
except their names, which I have replaced by single letters, at their
request.

“On the following Sunday, July 20th, at 8.15, there met in

Patrick’s rooms A, B, C, and D. Saturday being a Bank Holiday,
the plates were purchased on Friday evening by B, and kept by
him until the meeting. B produced his plates, unopened, and
after some preliminary table-turning and rapping, more
successful than at the previous meeting, it was decided to
proceed with the photographs. A carried the plate-box
unopened to the dark-room, and he and D sat closely on either
side of Patrick, and watched him open the box and load two
double dark-slides; they were satisfied there was no substitution
or trickery, or anything in the least degree suggestive of it. The
wrapper of the box was broken in full view of both, and Patrick
loaded the top four plates into two double dark-slides, which
were examined by A and D immediately before they were
loaded; they did not leave their sight from the moment of
examination until the photographs were taken. The camera was
also subjected to careful and minute examination, especially by
A, who removed the lens and examined both it and the interior
of the camera. The lens was then replaced, and the focal plane
shutter set in the open position, the exposures being made by
the simple expedient of withdrawing the shutter of the dark-
slide.
“At the request of C, before approaching the camera to focus it,
Patrick removed his coat, rolled up his sleeves, and was
carefully searched by him.
“It had been arranged that Patrick should take a photograph of
each of the four others present, under identical conditions. The
background was arranged, as before, of gowns hung over a
cupboard, but was made more complete. The subjects occupied
the same chair in succession; of the others, one stood by the
light switch, and the two others by the camera, to watch the
photographer. Patrick attended both to the camera and the flash
production. The exposures were made, as stated, by
withdrawing the shutter of the dark-slide; the focal plane
shutter was not touched throughout. The electric light was
therefore switched off for a few seconds while the shutter was
drawn and the flash being lighted. Sufficient light came through
the white window-curtains (9.30 p.m. Summer Time) to enable
those in the room plainly to see each other, and watch the
photographer’s movements. The four photographs were taken in
rapid succession.
“The slides were taken back into the dark-room, and developed
by A and Patrick in conjunction. B and C watched in turn, and D
also watched part of the time. One of the plates was quickly
observed to have an ‘extra’ developing on it. A bromide print
was again taken from the wet negative, and showed on the
photograph of D the head of an elderly man, besides a very fair
photograph of the sitter. The extra face was above D’s head,
and to his right. The “spirit” was bearded, and partly bald, with
a somewhat melancholy expression. There was a suggestion of
a white collar. On the left of the face and somewhat above it
was written in white on the black background what was
apparently a signature, with two final letters of a preceding
word. It was dubiously deciphered as ‘...ly S. Simmonds.’
Neither face, name, nor writing were recognised by any one,
either at the time or subsequently.
“The three other photographs were fair portraits, but showed no
abnormality.

“A third meeting was held in the same place at 8.15 p.m. on

Sunday, July 27th, when even more stringent conditions were
imposed on the photographer.
“The plates were bought on Saturday evening by D; other men
should have been present, but did not turn up at the arranged
time. D took the plates to his own rooms, where Patrick sealed
them for his own satisfaction. The box was kept locked up by D
till he brought them to the meeting on Sunday, and he did not
part with them till he gave them to E to take into the dark-room.
“At this meeting there were present A, B, C, D, E, F, and G,
besides the photographer.
“When all had arrived, E carried the plates to the dark-room. C
brought a dark-slide, which he had abstracted and kept since
the previous meeting. Before going into the dark-room Patrick,
again at the request of C and E, removed his coat, rolled up his
sleeves, and was searched, C even going to the length of
examining his socks for possible concealed plates or dark-slides.
“Patrick wished to load the slides himself, as they were rather
delicate. Accordingly neither slide nor plates were passed into
his hands until he was sitting in front of the ruby light, with E on
one side of him and C and F on the other. He broke the seals,
and in full view of these three loaded a single plate into
compartment No. 3 of the dark-slide. This was then immediately
taken from his hands again by E, and he and C locked it in a
drawer of the desk, upon which stood a reading-lamp, which
was never extinguished throughout all the subsequent
proceedings. C kept the key of the drawer, and passed it to E
when the slide was required.
“Some table-tilting was then carried out by all except C, who
remained at the desk and acted as secretary. The lights were all
put out except the reading-lamp he used, which was, as stated,
over the drawer where the dark-slide lay locked.
“After half an hour or so of moderate success with the table, E
and Patrick also dropped out, to take a flashlight photograph of
the group round the table. Patrick prepared the flash-powder,
and set up the camera—which had previously been examined—
by the side of the desk and lighted lamp. E again examined the
camera, inside and out, and when Patrick had focussed it
examined the view in the ground-glass screen. (The lights were
put up for a few minutes, to aid the focussing, etc.) When all
was ready, E received the key from C, unlocked the drawer, and
took out the dark-slide. He saw that it was undoubtedly placed
in the camera right way about, i.e., No. 3 compartment in use,
and the shutter withdrawn. When the table had commenced its
tilting again the flash was fired by Patrick. C took notes of the
movements of the table, and at the same time watched the
camera, which was in the full light of the reading-lamp
throughout. After the flash the shutter of the slide was replaced,
and on removal from the camera the slide immediately passed
again into the possession of E. Any substitution of plate or dark-
slide was thus rendered out of the question.
“The dark-slide was taken to the dark-room by E, and he and C
watched Patrick open it, remove the plate, and develop it. As
before, E kept the slide till everything was ready, and passed it
to Patrick in the full light of the ruby lamp, C checking the
number of the compartment in which the plate had been
loaded, and still remained (No. 3). On development, Patrick
pointed out that there was a hand at the top of the plate, which
could not belong to any of those at the table, and was pointing
with its index finger at one of the group. On fixing, it was
examined more closely, both by Patrick and the two others. All
three distinctly saw the image of a hand and wrist, pointing, the
forearm being draped. It was in fairly sharp focus, and
appeared, by its proportion, to be rather nearer the camera
than the centre of the table, above which it appeared to hang
suspended. A shadow cast by it was plainly seen, larger and less
sharply focussed, apparently on the back wall of the room. (A
picture on this wall had previously been removed, to eliminate
any reflection, and leave the background clear.) There was a
general appearance of drapery surrounding the group,
particularly at the sides; there was in this the suggestion of a
trunk to which the hand might belong. The appearance of the
picture was very startling, and Patrick suggested that as the
man at whom it should turn out to be pointing might suffer
considerable uneasiness on seeing it, it might be well to destroy
the plate without attempting to identify him. E and C, after a
minute’s thought, both agreed that this would be the wisest
course, and it was accordingly done. Patrick did not wish to feel
that he might be in any way responsible for causing anyone
uneasiness or harm, such as might well result from such a
 picture.[8] Accordingly the three returned to the other room,
and explained the situation to the others, who, though obviously
disappointed, did not condemn the course taken.
“This concludes the account of these first three meetings. We
wish to record that all through the meetings Patrick desired and
requested us to take all and any precautions we thought fit, to
satisfy ourselves that he introduced no trickery.
“In conclusion, we, the undersigned, declare this to be an
accurate account of the occurrences to the best of each man’s
individual knowledge. While not committing ourselves to any
statements as to our belief or disbelief in the genuineness of the
phenomena observed, we maintain that the greatest possible
care was taken to prevent any possibilities of trickery; and we
consider that, barring the possibility of Patrick having an
accomplice among us, the evidence should be accepted as proof
of the genuineness of the phenomena observed.”

This is followed by their seven signatures. E added afterwards a

paragraph of his own as to the interpretation of the word
“accomplice.” E was much the acutest observer and the most
obstinate sceptic of the seven: I think he suspected D of being in
some way my accomplice; some of the others suspected him of
being a medium. He certainly was not an accomplice—for I never
had one in the room; he may be a medium for aught I know—but I
should doubt it.
At the next meeting an eighth investigator appeared, and everybody
seemed to be suspecting everybody else, and not merely the
photographer. The plates were bought at a different shop, chosen by
lot, by a committee of four; and the packet was at once done up
with much red tape and green sealing-wax. When they had finished
I requested to be allowed to put my seal on it too, to assure myself
that they were not playing any tricks! My request was granted. I
now quote the report of the meeting:
 “The box of plates was produced by C, and the seals were found
to be intact. The box was taken into the dark-room by A, and a
plate-carrier—which had been previously examined by several of
those present—by B. The seals were broken, and a plate was
loaded in the presence of A, B, D, and E, who signed their
names on stamp-paper fixed to the back of the plate.
“In attempting to fit the slide into the camera, the plate was
accidentally exposed. It was discarded, and another plate
signed and loaded by A, C, E, and Patrick. C then locked the
plate away in a drawer, and kept the key until the slide was
required for the photograph.”
[Table-turning was then indulged in; A, C, E, and myself not
taking part. The usual type of answers was obtained from the
table; I omit this part of the report. During the table-tilting the
photograph was taken under precisely the same conditions as at
the last meeting.]
“The plate was developed by Patrick; A, C, and E watching. An
extra pair of eyes and the upper part of a nose developed,
apparently on the wall; they were brightly illuminated, from the
same position as the other figures. They were larger than those
of the other members of the group, and were over B’s head.
“We consider that this is a true account of what occurred.
Barring any very abstruse and elaborate explanation, it would
seem that the photograph is undoubtedly genuine.”

Then follow the signatures. As they made me sign the report on this
meeting, I had to see that it was worded rather carefully, particularly
the last paragraph; the report was true, so far as it went; and the
explanation of the result was rather elaborate; so I felt I could safely
sign it.
I did not hold another photographic séance, but being emboldened
by success, introduced at the next meeting “a medium from
London.” (As a matter of fact he came from Trinity, but I had
ascertained that nobody knew him, which was the important thing.)
After suitable preliminaries we all sat round a large table in semi-
darkness, holding hands. When the medium had arranged “the
balance of the circle” to his liking, he proceeded to go into a trance,
when queer things began to happen. A candlestick was seen to slide
along the mantelpiece and crash into the coal-box, taking a framed
photograph with it; sounds were heard from a small cupboard; the
window-curtains were parted; several people saw spirit forms and
eyes; and one was favoured with a spirit touch. The medium’s
Egyptian control, Nemetra, gave us wonderful accounts of life in
Memphis in the days of the Pharaohs—accounts which certainly
made up in picturesque detail for anything they lacked in historical
accuracy.
Unfortunately this meeting was not a complete success, as,
immediately the show was over, our ever-curious geologist E began
hunting about the floor, and discovered a small loop of fishing-line
(being a post-war fishing-line, the spirit forces had broken it). He
could not very well announce his find at the time, as the medium
was not yet roused from his trance, and the others were busy feeling
his pulse, fanning him and administering cold water!
By this time the results of the photographic séances had become
pretty generally known, and the undesired notoriety brought so
many requests to allow other visitors at the séances that it became
evident to me that the proceedings must terminate. So the next
morning, after seeing E, I told him and the others that the whole
thing had been a hoax, and that the photographs were frauds. I
should like to add that with one exception they took it extraordinarily
well, particularly when I explained what had been my object. They
were still quite in the dark about how the photographs had been
done, particularly when I told them that there was no accomplice
among them.
All the photographs were obtained by the general method of double
exposure and substitution, the substitution being effected at a
different point on each occasion; the methods used, or slight
variations of them, are all described in the section on “Methods of
Fraud.”
Now I maintain that the conditions imposed upon me were as strict,
or stricter, than any professional medium allows. If an amateur
photographer but little practised in sleight-of-hand can under such
conditions deceive intelligent observers—not once, but several times
over—how much easier will it not be for the professional spirit
photographer, who makes such frauds his business?

C.—Internal Evidence of Fraud

Since spiritualists claim that the presence of invisible spirits may be
detected by photography, it seems reasonable to inquire how far this
is compatible with established physical facts. If a plate is wrapped in
paper and submitted to “spirit influences”—whatever these may be—
never being exposed in a camera at all, and on development shows
faces or writing, I personally can only find one explanation—trickery.
But if a plate is duly exposed with camera and lens, and unseen
faces appear on development, the matter is not quite so simple. For
it is well recognised that the camera may record what is invisible to
the eye; invisible stars are detected by the photographic plate, and
anyone who has examined a nebula or comet through a telescope,
after seeing a photograph of the same object, realises this fact to his
disappointment. Similarly a can of hot water may be photographed,
by a long exposure, in a perfectly dark room; and another well-
known instance of a similar phenomenon is Sir Robert Ball’s story of
photographing some writing on the side of the “Great Eastern,” years
after it had been painted out and rendered invisible.
Light, as is well known, is now regarded as consisting of waves in
the ether. Ether waves are known to exist over a very large range of
wave-lengths; some are comparatively long waves, some are short.
The properties of these waves depend upon their wave-length; those
visible to our eyes, which we call “light rays,” form only a small
section of the complete scale; comparing them with sound waves
they correspond to approximately one octave of the whole musical
scale. Ether waves of greater or lesser wave-length than light, i.e.,
of lower or higher octaves, have very different properties. Radiant
heat and ultra-violet rays are the ether waves nearest in wave-length
and properties to light; X-rays and the waves responsible for wireless
telegraphy appear to be similar waves further removed along the
scale of wave-length.
Now in order to photograph an invisible object we require rays that
(a) affect a photographic plate; (b) are capable of refraction by a
lens; and (c) are invisible to the eye. The properties of the principal
known rays concerned may be summarised as follows:
Refracted by Lenses
Effect on Plates Visibility
Infra-red (heat)
v. slight Yes No
rays
Light rays affected Yes Yes
strongly affected
Ultra-violet rays Yes No
X-rays affected No No
It appears, then, that ultra-violet rays are suitable for our purpose;
infra-red rays, if present in an amount sufficient to affect a
photographic plate, would make themselves very evident as heat,
and may therefore be ruled out.
Ordinary daylight contains ultra-violet rays, as also does the light of
the arc lamp and magnesium flash; lamplight, gas-light, and the
ordinary electric light, are comparatively deficient in them. But are
we to assume that the spirit form is dependent on finding suitable
rays in the surrounding ether, or can it produce its own? Perhaps
some spiritualist will tell me. This is a point of some practical
importance in examining a reputed spirit photograph; for if the spirit
is self-luminous its features will be evenly illuminated and without
shadows, nor will it cast a shadow on the sitter or background, but
rather the reverse. If, on the other hand, the spirit is dependent on
the presence of ultra-violet rays from other sources, which it can
reflect, then the spirit in the photograph will appear to be
illuminated from the same point as the sitter,[9] and by absorption or
reflection of the ultra-violet actinic rays which would otherwise have
passed on, will cast a shadow on the background. Being a shadow
cast by the removal of the ultra-violet rays only, it will of course
appear as such in the photograph, but be invisible to the eye.
So if a spirit photograph is to be classed as possibly genuine, the
spirit may either appear self-luminous and cast no shadow, or may
appear to be illuminated from the same point as the sitter, and cast
a shadow on the background, if the latter be of a suitable nature to
show it. But on examining a collection of spirit photographs taken by
various professional mediums, we find that as often as not the spirit
and sitter are lighted from opposite sides; or that a spirit face with a
well-marked shadow on one cheek throws no shadow on the
background. If our reasoning be correct, we can at once write such
productions down as frauds. The photographs I produced at my
Cambridge séances show both these faults; two of them have the
spirits lighted from the opposite side to the sitter, and one has the
spirit lighted from the correct side but throwing no shadow, whereas
the sitters throw clear shadows on the wall behind. In the other
photograph I managed to get both the lighting and the shadow of
the spirit correct; but in order to get the shadow I had to
photograph the background with the “spirit”; hence when the sitters
were photographed on the same plate there was a double
background, which necessitated a rapid destruction of the plate!
Of course the average medium does not consider these points at all;
his sitters are usually satisfied with anything they can get, so why
should he worry? But an intelligent observer examining a number of
spirit photographs with regard to these points will quickly satisfy
himself that the majority of them can only be frauds.[10]
There are a number of other points by which a spirit photograph
may betray its method of production without reference to the
conditions under which it was taken. Many spirit extras are simply
copies of existing photographs, which are usually camouflaged in
some way. Draperies may be substituted for the hair, or the features
slightly retouched. A common method is to reverse the original
photograph, right for left; a number of Hope’s productions were
recently published in a monthly magazine, and alongside them life
portraits of the “spirits,” the letterpress emphasising that, though
undoubtedly the same face, they were different photographs. On
examination with a mirror, however, the photographs were seen to
be identical, and careful measurement of the faces showed the
proportions to be exact. In the photographs more recently published
by Mr. Bush, who laid a trap for Hope into which the latter appears
to have fallen, the spirit was not reversed, nor was even the rather
peculiar attitude of the head in the original photograph altered. A
little spirit drapery was added round the face, and the whole thrown
slightly out of focus; it is really a most clumsy piece of work, and
should deceive no one.
In some spirit photographs produced by double exposure there is a
double background, as occurred in my own photograph referred to
above. There may be either two different backgrounds, or a double
outline of the same background; in either case the “spirit’s
background” is usually fainter than the “sitter’s background,” and
shows through the darker parts of the sitter. Sometimes attempts
are made to retouch these appearances on the negative, and many
spirit photographs show clumsy brush or pencil work, which must
immediately stamp them as frauds.
Attempts are sometimes made to obliterate other tell-tale marks,
such as a piece of a spirit’s hat or collar, which has accidentally got
on to the plate. Other mediums, however, are less particular,
especially in America, and produce their spirits with ordinary hats,
collars and ties. But as a rule only spirit robes are permitted,
apparently made of butter muslin not quite in focus. Hands are often
present: I have seen a case in which the position of a spirit hand
would have necessitated a many-jointed arm about four feet long;
but perhaps spirit arms are like this. One spirit extra I have seen has
two hands, but both appear to be left hands—evidently a left-
handed spirit.
Frequently, again, careful examination shows that spirit extras are
not photographs at all, but resemble wash drawings. This gives the
clue to their origin, for several of the methods described in a
preceding section produce a result of this kind. It has been several
times pointed out that spirit extras in some cases show the
characteristic dots produced by the half-tone newspaper illustration
process; if the medium cannot obtain a real photograph of the
required spirit, he has to copy a newspaper reproduction. If he is
clever, he can eliminate these process marks by printing in his spirit
slightly out of focus; but very often he does not take the trouble.
In many, perhaps in the majority, of spirit photographs produced by
professional or semi-professional mediums, a critical observer with
practical photographic experience can point out some such definite
evidence of fraudulent manipulation. In many other cases, where no
one particular point can be singled out as indicative of fraud, minor
points of suspicion are noticeable, which taken together leave little
doubt of the nature of the picture. But photographs can be prepared
by purely mechanical means, especially if no kind of test conditions
are employed, which will contain no internal evidence whatever of
manipulation. By carefully combining enlarged positives, for
instance, and re-photographing the whole, results can be produced
which will defy the most critical examination. But such photographs
are seldom produced, even when the medium is given practically a
free hand.
 IV.—Spirit Photographs Obtained by
Amateurs
(C. Vincent Patrick)
Probably most people have heard, but seldom at first hand, of
unexpected ghosts appearing on plates or films exposed by amateur
photographers. On the rare occasions when such accounts can be
traced to their source, one usually finds that there is some simple
and evident explanation. Streaks and splashes of light on the plates
are comparatively common, and are usually the result of the camera,
slides, or dark-room not being light-tight; very strange results are
sometimes produced in this way. I was once puzzled by a
photograph which showed an arch, like a rainbow, across the sky,
when it was quite certain that there had been no rainbow in the sky
when the photograph was taken. When the result was repeated a
few days later, the camera quickly came under suspicion, and was
found to have developed a minute pinhole in the bellows. This was
sealed up, and the rainbow did not reappear. Many unexplained
markings on plates are certainly caused in this or similar ways; but
only under very favourable circumstances could an extra face on the
plate be so produced. Sometimes unexpected results are caused by
an accidental second exposure; but the nature of such a photograph
will quickly be apparent. The use of old glass plates may sometimes
be responsible for similar results, as has been already explained. But
authenticated cases of the appearance of unseen faces in
photographs taken in the absence of a professional medium, and
which do not show an obvious explanation, are few and far between.
The classical example is that of the Combermere photograph, which
was published in the Journal of the S.P.R., and aroused much
discussion and criticism.
A Miss Corbet took a photograph of the library of Combermere
Abbey, Cheshire, on December 5th, 1891. She was alone at the time,
and left the camera during the exposure, as it was a long one. She
kept a note-book with records of her photographs, which afterwards
showed that an exposure of one hour had been given, namely from
2 to 3 p.m. Unfortunately she did not develop the photograph till
eight months later, and was then amazed to find a figure occupying
a chair in a prominent position in the photograph. The figure was
faint and transparent, the legs being quite invisible; the features
were not recognisable; but the presence of a head, shoulders and
arm was fairly plain. Inquiries were made, and it was found that not
only was the chair in question the one Lord Combermere had been
wont to occupy, but that he had died a few days before the
photograph was taken, and was actually being buried some two
miles from the Abbey at the hour at which the photograph was
taken. The photograph was naturally shown to the dead nobleman’s
relatives, some of whom professed to recognise it as Lord
Combermere. It was further pointed out that he had lost the use of
his legs in an accident some three weeks before his death, and that
the spirit figure was correspondingly legless!
The most important contribution to the discussion which followed
was made by Sir William Barrett, who demonstrated that the result
could be duplicated by taking a several minutes’ exposure of a chair,
in which someone was seated for a part of the time. The sitter would
naturally not keep quite still; hence the outlines would be blurred
and the features indistinct. Sir William published a photograph which
he had obtained in this way, reproducing the features of the
Combermere photograph, even to the leglessness. He suggested
that someone, possibly one of the four men-servants in the Abbey,
had entered the library during the prolonged exposure. He had sat
down in the chair for a minute or so, when, noticing the camera, he
beat a retreat. The photograph showed double outlines to all the
sharp edges, indicating that the camera had been moved slightly
during the exposure, and suggesting that someone had entered the
room and jarred it. As it was eight months after the event that the
photograph was developed, it was impossible to ascertain whether
anyone did actually so enter the room. In any case it was a
remarkable coincidence, but there is no proof of it being anything
more.
A somewhat similar case is recorded by Podmore. The photograph
was being taken, this time, in a chapel. On development a faint face
was seen framed in a panel. This was described as being the
likeness of a friend of the photographer’s who had recently died—“a
handsome, melancholy lad of eighteen.” Another critic thought that
the face was that “of a woman of thirty”; it must have been very
indistinct. It may well have been caused in the same manner that
was suggested for the Combermere photograph; a visitor to the
chapel standing in the field of the camera for some moments,
probably not realising that an exposure was in progress.
Several accounts have been given by amateurs of seeing spirit faces
develop, only to disappear again on fixing; one such is published in
Vol. VII. of the J.S.P.R. These are evidently of a subjective nature,
the finished negative showing no evidence of any abnormality. If any
reader of this article knows of any case where an “extra” has been
obtained in the absence of a professional medium, and where the
plate can be produced, I should be very grateful for particulars.
Experiments have on several occasions been made by amateurs,
deliberately trying for spirit extras, and exposing scores of plates,
usually without success. The unsuccessful attempts of Russell,
Beattie, Dr. Williams, and more recently Dr. Pierce, have already
been alluded to. Experiments of rather a different nature have been
carried out by a Frenchman, Dr. Baraduc. His most interesting—if
somewhat gruesome—result was a series of photographs taken over
the death-bed of his wife, at the time of, and for some hours after,
death. The negatives showed globes of light floating over the bed,
which gradually increased in size and brightness, and coalesced in
the later photographs. The circumstances certainly seem to exclude
fraud, and it is very difficult to understand how the progressive
series of photographs could have been obtained by accidental
means, such as a pinhole in the camera. His results are very
interesting, but need repeating by other experimenters; in any case,
they have absolutely nothing in common with the conventional spirit
photographs which show faces and figures.
 V.—The Fairy Photographs
(C. Vincent Patrick)
The so-called “Fairy Photographs” recently published by Sir Arthur
Conan Doyle and Mr. E. L. Gardner do not strictly come under the
heading of “spirit photographs,” but may not inappropriately be
considered here. We have no evidence of the conditions under which
they were taken; as Sir Arthur explains, such “rare results must be
obtained when and how they can.” We have therefore to learn what
we can from an examination of the photographs, or of their
reproductions. At first sight they look like genuine untouched
photographs; their general appearance is excellent, and if frauds,
they are certainly good ones. On examining them more carefully,
however, a considerable number of points are found requiring
explanation. Some of these have no doubt been noticed by different
observers; the principal criticisms of the different photographs are
these.
“Iris and the Dancing Gnome” shows some very strange lighting.
Examining Iris’s hat, we find the strongest light is falling, probably
through a gap in the trees, from above and a little to the right; the
shadow behind her arm, and the lighting of the fingers, confirm this.
The gnome stepping up on to Iris’s knee should therefore cast a
shadow upon her white dress, below and to the left; but the
photograph shows no trace of any such shadow. On the other hand,
the gnome is lighted mainly from the left; this is plainly shown on
the conical cap and the right upper arm. Apart from these
discrepancies, which alone are quite sufficiently damning, several
other grounds for suspicion are evident. The whole photograph is
much too carefully arranged to be the snapshot it is represented as
being. The black legs of the gnome are contrasted against the white
dress of the girl; the lighter body, face and wings are outlined
against the shadows under the trees; the dark cap is brought with
one edge against a wing, the better to show it up, while the other
edge catches the light. A snapshot would indeed be fortunate in
securing such an admirable arrangement! The same thing is very
noticeable in the other three published photographs; the pictorial
arrangement of the figures and background is much too good to be
the result of chance, and suggests careful posing.
This gnome photograph was taken under the shade of trees, we are
told, at four o’clock on a September afternoon which was not sunny;
an exposure of 1⁄50th of a second was given on “Imperial Rapid”
plates, using a “Midg” quarter-plate camera. With the largest stop in
this camera an exposure of at least ten times that stated, i.e., 1⁄5th
of a second, would be needed to give a fair negative under these
conditions; 1⁄2 to 1 second would probably be more correct. The
photograph in question certainly shows signs of under-exposure; but
under the conditions stated one would expect little more than a
silhouette of the white dress and of the sky showing through the
trees. Something is evidently wrong here.
The gnome’s proportions are certainly not human, as are the fairies’
in the other photographs; he rather resembles the familiar “Brownie”
of the Kodak advertisements. Though stepping up onto the girl’s
knee, he is noticeably looking away from her, and at the camera,
which is very unnatural and likely to cause him a tumble! Criticism
has been directed against the girl’s hand, but this is quite a common
photographic distortion of a hand held rather near the camera. In
my copy, however, the elbow appears rather peculiar.
The other points, taken together, can leave no possible doubt that
the photograph is a fake. It could have been produced by making a
positive enlargement from the negative of Iris on one of the bromide
papers specially prepared for working up. The gnome would then be
sketched on this—he certainly resembles a sketch more than a
photograph—and the whole would then be re-photographed on to a
quarter-plate. No doubt an entirely satisfactory result would not be
secured at the first attempt; in fact, Mr. Gardner tells us that “other
photographs were attempted, but proved partial failures, and plates
were not kept.” Surely such extraordinary photographs, even if
partial failures, would be kept—if they did not show something that
was not intended! We have known plates to be destroyed on other
similar occasions, and for similar reasons.
“Alice and the Fairies” is of a rather different nature. The lighting of
the fairies is badly wrong; they are brightly illuminated from a point
behind the camera, whereas Alice is less brightly illuminated, and
from the left-hand side. Sir Arthur, in his article, points out that this
is accounted for by the “fairy psychoplasm” having a “faint
luminosity of its own.” To appear brighter than the sitter,
photographed by 1⁄50th of a second exposure at three o’clock on a
sunny July afternoon, the fairies would have to resemble in
luminosity a battery of arc lights! The photograph appears to have
been produced by pasting the “fairies” on to an enlargement of the
original photograph of Alice, and then re-photographing the whole.
The fairies could be obtained by taking posed photographs of
children suitably dressed; these would then be carefully cut out from
their backgrounds and pasted on to the original enlargement. The
points of internal evidence on which this statement is based are as
follows:
1. The very sharp (cut) outlines of all parts of the fairies. This is
particularly noticeable in the outline of the dress and hair of the third
fairy (counting from the left); compare this with the soft outline of
Alice’s hair, against a similar background.
2. The same fairy’s forearm is much brighter than Alice’s wrist, at the
point where it crosses between it and the camera. Assuming that
both were equally white, and lighted from the same source, the one
further from the camera would normally photograph a little the
lighter.
3. Fairies two and four appear to be photographs of the same model,
the wings being exchanged for the pipe. Note the similarity of the
attitude of the legs, and of the shape of the tail of drapery hanging
down behind.
4. With the exception of one foot of each of these fairies, which
appears somewhat unnaturally amputated, every part of the fairy
figures is in front of the sitter and background. This applies to all
four photographs, and is of the utmost importance; superimposing
the fairies on the original photograph in the manner described must
of course produce this effect.
5. One would have expected to see some blurring due to movement,
in the fairies’ wings and feet at any rate, with a 1⁄50th of a second
exposure at a distance of four feet. None is visible in the
reproduction.
The two more recently published photographs are very similar to
“Alice and the Fairies,” and the same general criticisms apply. “Alice
and the Leaping Fairy” again shows the fairy illuminated from a point
behind the camera, whereas Alice is illuminated from the right side.
(Note that her right cheek, facing the camera, is in shadow.) Fairy
shows no movement-blurring, and comparison with instantaneous
photographs of jumpers shows the attitude to be most unusual. On
tilting the photograph a little to the left, the fairy appears to have
been posed kneeling on the left knee, the support being afterwards
cut away, and the cut-out figure applied to the enlargement of Alice,
in a slightly different vertical axis.
“Iris and Fairy with Harebells” shows similar features. Notice, again,
the different lighting of fairy and Iris; the hard outline of fairy’s hair,
so unlike Iris’s in the same print; and the careful way the fairy has
been placed to secure a well-balanced picture—scarcely a random
snapshot! The harebells seem too large in comparison with the
hedge-leaves at the same distance from the camera. They may be
the result of combining yet a third photograph; or the actual
harebells may have been placed on the enlargement and re-
photographed with it.
An artist to whom I have shown this photograph, together with the
full-length photographs of “Iris” published with the earlier article in
the Strand Magazine, is of opinion that the fairy has the same figure
and features as Iris, and, in fact, may very well be a photograph of
Iris herself, attired in a bathing dress and some butter muslin, and
with the addition of wings! The photographs of Iris show a rather
characteristic poise of the head, which is also seen in the fairy. This
is only a suggestion, however; the photographs are too small for
certain identification. In any case, the fairy figure is certainly of
human proportions.
These photographs have attracted a good deal of attention, and
seem to have been accepted as genuine in some quarters. No doubt
much reliance has been placed on the statement of one experienced
photographer, Mr. Snelling, that they show no evidence of
manipulation, disregarding the adverse criticisms of several other
photographers to whom they were shown. I consider that there is
not the slightest doubt that they are fakes, simply on the internal
evidence they provide, and I have endeavoured to explain the
principal points on which this opinion is based.
 VI.—The Reliability of Witnesses
(W. Whately Smith)
The reliability of witnesses is a crucial question in the study of
psychical phenomena and has for long been a bone of contention
between spiritualists and their critics. If honesty, care, and
intelligence alone sufficed to make a man’s testimony reliable the
whole range of spiritualistic phenomena, including spirit
photography, might long ago have been taken as proved beyond all
possibility of doubt. But this is very far from being the case, and
although it is never pleasant to express flat disbelief of the accuracy
of people’s statements, the Psalmist’s dictum that “all men are liars”
should be graven on the heart of every psychical researcher,
especially in the case of those who attempt to investigate “physical”
phenomena.[11]
I do not propose to repeat the obvious platitudes about the ease
with which conjurers can deceive their audiences, but I should like
to emphasise the fact that such differences as exist between the
circumstances in which conjurers and mediums work are uniformly in
favour of the latter as regards the minor manipulations necessary for
the production of photographic phenomena. (One is not, of course,
concerned with elaborate “stage effects,” but rather with small
matters like the substitution of one plate for another or the
distraction of the sitter’s attention while the required extra is
impressed upon the plate.) The conjurer’s audience knows that it is a
trick; the medium’s does not. Even the most hardened sceptic will
probably have a lingering doubt in his mind as to whether there may
not possibly be “something in it” after all. This is all to the medium’s
advantage, and it must be remembered that not only does he work
for much of his time under lighting conditions which are peculiarly
favourable to fraudulent manipulation, but also that the great
majority of his sitters start with a considerable prepossession to the
effect that they are encountering something inexplicable.
But these observations must, I suppose, have occurred to all who
have considered such matters at all impartially, and however
relevant they may be they will never by themselves prevail against
what we call “the evidence of our senses.” No amount of general
considerations of this kind will deter the credulous from accepting
the prima facie indications of a “successful” séance. The only hope
of preserving the public from the depredations of these swindlers is
to show that the “evidence of the senses” is not worth twopence
unless backed by special knowledge of the relevant technique.
One would think that anyone who reads Mr. Patrick’s admirable
account of fraudulent methods and of his experiments in their
application will feel chary of claiming that he has wholly eliminated
the possibility of fraud from any photographic séance which he has
attended. But there may be some who will still say: “No doubt these
fraudulent methods can be and have been employed, no doubt
many people would allow a medium to substitute plates under their
very noses, or to touch them. But when I went to such-and-such a
medium I am certain that the plates were never out of my
possession, that he never had a chance of touching them....” and so
forth.
Of course, some of the methods described by Mr. Patrick do not
involve touching the plates at all. It would not be at all impossible
for an artist in such work to allow a sitter to use his own plates,
camera, slides, dishes, and chemicals in his own studio and dark-
room, to load, unload, and develop the plates himself without their
ever being touched by the “medium” and yet to produce a perfectly
good extra.
But I will let that pass and confine myself to the question of whether
the kind of positive statement outlined above is really worth
anything at all. This question was answered once and for all in the
emphatic negative by the classical experiments of the late Mr. S. J.
Davey in “Slate-writing,” which are fully described in the Proceedings
of the Society for Psychical Research, vols. iv. and viii.
These experiments are not nearly so widely known as they deserve
to be, but it is not too much to say that no one who has not read,
marked, learned, and inwardly digested them is competent so much
as to begin to talk about the genuineness of spirit photography;
unless, of course, he happens to have acquired a knowledge of trick
methods and the scope of deception by other means—such as Mr.
Patrick adopted in his experimental work!
Very briefly, the story was as follows: Mr. Davey was an amateur
conjurer of some skill who set himself to imitate by trickery the
performances of Slade, Eglington, and other exponents of “slate-
writing” phenomena. In this he succeeded to admiration—so much
so that certain spiritualists characteristically insisted that he must be
a very powerful “medium”! He scrupulously denied himself the
advantage of claiming his results as supernormal, but in spite of this
found no difficulty in imposing on his sitters. The latter were
encouraged to take every possible precaution against trickery and
were instructed to write the most careful reports of what occurred.
A number of reports were thus obtained from men and women of
unquestionable intelligence and acumen which, if they had been
even approximately accurate, would have established the
supernormality of Mr. Davey’s phenomena beyond any peradventure.
But comparison of their reports with the known and recorded
procedure which actually took place showed the most astonishing
discrepancies. Omissions and distortions of the first importance were
abundant and the experiments proved to the hilt that, for
phenomena of this kind, the reports of untrained witnesses are, in
general, not worth the paper they are written on.
I wish that space permitted me to quote, in parallel columns, some
of these Davey reports and some of those given by witnesses of
photographic séances so that my readers could see how very similar
the circumstances are.
But I must content myself with pointing out that whereas in the one
case everything turned on whether the “medium” had any chance of
substituting or tampering with slates, so in the other it is a matter of
whether there has been any chance of substituting or tampering
with plates. The reports of intelligent witnesses proved worthless in
the one case, and it seems reasonable to suppose that they are no
more valuable in the other.
So, to anyone who thinks that in the mouth of two or three
witnesses the genuineness of spirit photographs shall be established,
I would say, “Go home and invest a few shillings in the Proceedings
of the Society for Psychical Research, vols. iv. and viii.—it will be
more profitable than the same amount laid out in photographic
séances—and when you have carefully read their account of the
Davey experiments in conjunction with Mr. Patrick’s paper, see
whether your confidence in spirit photographs is as strong as ever!”
I have drawn attention to these experiments of Mr. Davey elsewhere
and I am sorry to be obliged to insist on their importance again. But
until people learn that the reports of uninstructed observers—
however acute in other respects—are utterly unreliable, the
fraudulent medium will flourish and the unsuspecting public will be
robbed and deceived.
 VII.—The Value of Recognition
(W. Whately Smith)
Believers in spirit photographs generally consider that they are
playing their trump card when they point out that thousands of
“extras” have been definitely recognised by sitters as portraits of
their deceased friends or relatives. But this card, impressive as it
looks, will not really take the trick. If it could be shown (i.) that a
given “extra” was unmistakably recognisable as a portrait of a
deceased—or even of a living—person, and (ii.) that the medium
concerned could not possibly have obtained a likeness of that person
to work from, then we should be obliged to attach great weight to
this factor, even if the conditions were not otherwise such as to
exclude fraud. For such a result could not be fraudulently produced.
But in spite of the perfectly honest assertions of many investigators,
it seems very doubtful whether this state of affairs has ever been
realised.
There are two ways in which evidence based on recognition may be
defective.
First, the recognition may be perfectly well founded, but the “extra”
may have been derived from an existing photograph of the
deceased; second, and more frequently, the recognition is illusory
and exists only in the sitter’s imagination.
As regards the first of these points, it should be remembered that
most people are photographed at one time or another, some of them
frequently, and that it is not very difficult to obtain a photograph of a
given person if one goes about it in the right way. A spirit
photographer with an extensive clientèle will find it well worth his
while to take the necessary steps to secure photographs appropriate
to at any rate his more regular sitters, from whom, in the course of a