DEEP LEARNING-BASED QUALITY ASSESSMENT OF 3D POINT CLOUDS WITHOUT

REFERENCE

Aladine Chetouani1 ,2 Maurice Quach2 Giuseppe Valenzise2 Frédéric Dufaux2

2021 IEEE International Conference on Multimedia & Expo Workshops (ICMEW) | 978-1-6654-4989-2/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICMEW53276.2021.9455967

1
Laboratoire PRISME, Université d’Orléans, Orléans, France
2
L2S, Centrale Supélec, Université Paris-Saclay, Gif-sur-Yvette, France

ABSTRACT been an increasing interest in the multimedia research com-

munity towards the definition of accurate quality metrics for
Point cloud (PC) quality assessment is of fundamental im- point clouds. These efforts have been also partially motivated
portance to enable the efficient processing, coding and trans- by the concurrent standardization activities for point cloud
mission of 3D data for virtual/augmented reality, autonomous coding in MPEG [1], as well as due to the need of relevant loss
driving, cultural heritage, etc. The quality metrics proposed functions in learning-based PC compression methods [2, 3].
so far aim at quantifying the distortion in the PC geome-
try and/or attributes with respect to a reference pristine point Different point cloud objective metrics have been pro-
cloud, using simple features extracted by the points. In this posed in the literature. Point-to-point metrics were among
work, we target instead a blind (no-reference) scenario in the first to be considered, and compute geometry distance be-
which the original point cloud is not available. In addition, we tween corresponding points in the original and distorted PC.
learn features from data using deep neural networks. Given On the other hand, the point-to-plane metric is an extension
the limited availability of subjectively annotated datasets of of the previous metric and consists in projecting the point-to-
corrupted point clouds, and the consequent difficulty to learn point error vector along the local normal [4]. Starting from
in an end-to-end fashion PC quality features, in this work we these approaches, several point-based metrics have been then
use instead a two-step procedure. First, we extract from lo- developed. In [5], the authors proposed to estimate the geo-
cal patches three relevant low-level features which have been metrical error between two PCs (i.e., reference PC and its dis-
commonly used in other PC quality metrics, i.e., geometric torted version) by measuring the angular similarity between
distance, local curvature and luminance values. Afterwards, tangent planes. In [6], the authors proposed a metric called
we employ a deep neural network to learn, from these low- PC-MSDM by extending the well-known SSIM metric [7],
level features, a mapping to the PC ground truth mean opinion widely used for 2D images, to PC, by considering features
score. Our results on two state-of-the-art PC quality datasets including local mean curvature as they previously did for 3D
show the potential of the proposed approach. The code is meshes [8]. The authors proposed later a metric called PCQM
available here : https://github.com/achetouani/ that integrates color information [9]. In [10], the authors pro-
NR-CNN-3D-PC posed a new approach that focuses more on the distribution
of the data. They introduced a new type of correspondence
Index Terms— Point clouds, quality assessment, deep from point to distribution characterized using the well-known
learning, no-reference Mahalanobis distance. In [11], the authors proposed a color-
focused metric that integrates geometry information. In [12],
1. INTRODUCTION the authors adapted also the SSIM metric for point clouds us-
ing a number of features, while in [13] the authors improved
3D point clouds (PC) are becoming and increasing popular the point cloud PSNR metric. The authors of [14] propose
data format for applications such as virtual and augmented re- a different approach, by studying the effectiveness of learned
ality, autonomous vehicles, cultural heritage, immersive com- features for the prediction of PC quality, obtaining some en-
munication, etc. PCs can undergo different types of distor- couraging results. Interesting approaches for 3D meshes have
tion, due to acquisition, compression, transmission or render- also been proposed [15].
ing. Predicting and quantifying the perceptual impact of this
In this paper, we present a deep learning-based frame-
distortion is therefore of paramount importance for the de-
work that efficiently predicts the quality of PCs without refer-
ployment of this kind of data. In the past few years, there has
ence. The proposed method consists first in extracting patches
Funded by the ANR ReVeRy national fund (REVERY ANR-17-CE23- around a set of points from the PC. The regions delimited
0020). by the extracted patches are then characterized by three at-

978-1-6654-4989-2/21/$31.00 ©2021 IEEE

Authorized licensed use limited to: University Roma Tre AREA SCIENTIFICO TECNOLOGICA. Downloaded on December 11,2023 at 04:20:04 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. General framework of the proposed method.

tributes: geometric distance, mean curvature and gray-level. distorted PC, we first randomly select a set of N points. The
The resulted patches (i.e., one per attribute) are stacked and latter are here considered as reference points from which the
fed as input to a Convolutional Neural Network (CNN) model quality is predicted. Then, we delimit a region around each
to predict their quality. The global quality index is finally point Ni by finding its K nearest neighbors. Next, the neigh-
given by aggregating the predicted patch quality indexes. We bors of each point Ni are characterized by three attributes:
assess the performance of our method on two datasets, includ- geometric distance, mean curvature and gray-level. The ge-
ing a large dataset more suited to deep models. The high cor- ometric distance D describes the spatial distribution of the
relations obtained showed the potential of the approach. The neighbors and it is computed as follows:
main contribution of this paper is the use of classical CNN
 
models to predict the quality of 3D PC through extracted fea- D Kji , Ni =
tures.  2  2  2
The rest of this paper is structured as follows: Section 2 xKji − xNi + yKji − yNi + zKji − zNi , (1)
describes the proposed framework. Experimental results are
presented in Section 3 and the conclusion is given in Sec-  
where D Kji , Ni represents the Euclidean distance be-
tion 4. s
tween Ni and its j th nearest neighbor Kji . {x, y, z}Kji and
2. PROPOSED METHOD {x, y, z}Ni are the spatial coordinates of the j th neighbor
and the point Ni , respectively.
As illustrated by Fig. 1, the framework proposed for estimat-
ing the quality of 3D point clouds without reference is based The mean curvature allows to characterize the shape vari-
on 2 main steps: 1) feature extraction that aims to charac- ations and it is computed for each point of the delimited re-
terize the PC through a set of features as patches from se- gion, including the point Ni . The latter is estimated through
lected points and, 2) quality prediction that aims to estimate a quadric fitting. The grey level gives information more re-
the quality of the distorted PC via a CNN model. Both steps lated to the perceived rendering and it derived from the color
are described in details in this section. information (i.e. RGB to grey level). The features obtained
for each attribute are encapsulated into a patch and the three
2.1. Feature extraction as patches
resulted patches are then stacked to form a new patch SPi of
size P ×P ×3 where three represents the number of attributes
considered.

2.2. Quality prediction

Fig. 2 gives a detailed overview of the steps applied to predict
the Global Quality Index (GQI) of the PC from the stacked
patches SPi .

2.2.1. CNN models and patch quality indexes

For each stacked patch SPi , we first predict its Patch Quality
Index (i.e. P QIi ) through a CNN model. To do so, we com-
pared here 5 pre-trained models that were widely used for
Fig. 2. Random point selection and patch extraction. classification tasks and some of them were successfully em-
ployed in several studies related to quality assessment as well
Fig. 2 details the feature extraction process. From a given [16]. It is worth noting that these models were considered in

Authorized licensed use limited to: University Roma Tre AREA SCIENTIFICO TECNOLOGICA. Downloaded on December 11,2023 at 04:20:04 UTC from IEEE Xplore. Restrictions apply.
 order to compare the performance of different architectures where N is the number of stacked patches (i.e. the number
and analyze the impact of the depth. of selected points).

• AlexNet [17]: This model is one of the pioneering It is worth noting that a plethora of handcrafted and deep-
models proposed by Alex Krizhevsky. It highlighted based methods already aggregated local predicted scores to
the relevance of using CNN models for classification derive a global quality index like SSIM [7], VDP [24] and the
tasks. The authors brought out three main points: the following deep-based methods as well [20, 25]. This strategy
use of the Relu (Rectified Linear Units) function, the was also applied for 3D content as done in [6, 8].
exploitation of the dropout to prevent the over-fitting
and overlap during the pooling step.

• VGG [18]: This model was developed by the Oxford

Visual Geometry Group in 2014. It is characterized by
applying a max pooling layer after a succession of con-
volution layers. Here, we used V GG16 and V GG19 Fig. 3. Prediction of the Global Quality Index (GQI) from the
which are composed of 16 and 19 layers, respectively. stacked patches SPi .

• ResNet [19]: This model was introduced in 2015 and it

stands out by the integration of a residual module. The 3. EXPERIMENTAL RESULTS
authors proposed to reformulate the output (H(x)=F(x))
of each series of Conv-ReLu-Conv by adding the in- In this section, we evaluate the capacity of our method to pre-
put x as information (H(x) = F(x)+x). ResN et18 and dict subjective scores. For that, we first present the datasets
ResN et50 are used in this study. used and describe the protocol applied to train and test our
CNN models. After defining the performance evaluation cri-
The above-models were adapted to match their inputs and teria used, we discuss the results obtained by each considered
outputs to our context. We changed the size of their input lay- CNN model and compare the best results with state-of-the-art
ers and replaced their FC layers with three FC layers of size PC metrics.
64, 64 and 1, respectively. The two first FC layers are fol-
lowed by a ReLu layer, while the last FC layer is preceded by 3.1. Datasets and experimental setup
a a dropout layer with a probability fixed to 0.5 and aims to
predict the PQIs of the stacked patches (i.e. regression layer). The proposed method is evaluated using two recent 3D point
We finally fine-tuned each model to adapt their weights to cloud datasets: sjtu [26] and ICIP20 [27].
our context. It is worth noting that the target of each stacked
• sjtu is composed of 9 point clouds from which 378 de-
patch was the subjective score of the whole distorted PC as
graded versions were derived (i.e. 42 distorted PC per
commonly employed to estimate the quality of several mul-
reference PC) through 6 degradation types (OT: Octree-
timedia content for 2D images [20, 21] as well as for stereo
based compression, CN: Color Noise, DS: Downscal-
images [22] and 3D meshes [23].
ing, D+C: Downscaling and Color noise, D+G: Down-
During the training step, the stochastic gradient descent
scaling and Geometry Gaussian noise, GGN: Geometry
optimization algorithm was used with the Mean Squared Er-
Gaussian noise and, C+G: Color noise and Geometry
ror (MSE) as loss function. The batch size, learning rate and
Gaussian noise).
the momentum were set to 64, 0.0001 and 0.9, respectively.
We fixed the number of epochs to 30 with a validation each • ICIP20 is composed of 6 commonly used point clouds
5000 iterations. At each epoch, the training data was shuffled from which 90 degraded versions were derived (i.e. 15
and the resulted model was saved. The model that provided distorted PC per reference PC) through 3 types of com-
the best results was finally retained. pression: V-PCC, G-PCC with triangle soup coding and
G-PCC with octree coding. Each reference point cloud
2.2.2. Global Quality Index was compressed using five different levels.
Figure 3 illustrates how to obtain a Global Quality Index Our method is evaluated by decomposing each dataset
(GQI) that reflects the quality of the whole distorted PC by into F folds. At each fold, one reference PC and its distorted
aggregating the predicted PQIs. To do so, we aggregate the versions are used to test the model, while the rest are used to
obtained PQIs by computing the average scores as follows: train the model. The number of folds is here equal to 7 for
N
sjtu and and 6 for ICIP20.
1  In this study, the number of selected points (N ) was set to
GQI = P QIi , (2)
N i=1 1000, while the number of neighbors (K) and the size of the

Authorized licensed use limited to: University Roma Tre AREA SCIENTIFICO TECNOLOGICA. Downloaded on December 11,2023 at 04:20:04 UTC from IEEE Xplore. Restrictions apply.
 patch (P ) were fixed to 1023 and 32x32, respectively. The • The architecture seems to have an impact since VGG-
patch size was set following studies where its impact on the based models give the best correlations with an im-
performance was analyzed [20, 25]. provement mean PCC gain around 5%. Moreover,
unlike for classification tasks the integration of resid-
3.2. Performance evaluation ual modules doesn’t improve the performance. The
p-values between the VGG-based and ResNet-based
Two evaluation criteria commonly used to evaluate the perfor- models are smaller than 0.05, indicating a statistically
mance of quality metrics are adopted here: 1) Pearson Corre- significant difference between the two.
lation Coefficient (PCC) and 2) Spearman Rank-Order Coef-
ficient Correlation (SROCC). Both vary between 0 and 1 in Based on these results, it seems that the architecture of the
absolute value, with 1 being the best performance. These cor- considered models has more impact on the performance than
relations are computed for each dataset over each fold and the depth.
the mean correlations are then reported as results. It is worth
noting that the same procedure was applied to the compared 3.4. Comparison with state-of-the-art metrics
state-of-the-art metrics. Our method is here compared with state-of-the-art metrics.
More precisely, we consider po2point-based and po2plane-
3.3. Model comparison based metrics pooled with MSE, PSNR and Hausdorff. We
also considered recent metrics po2dist [10] (i.e. point to dis-
tribution) pooled with MSE and PSNR as well as PCQM
Model PCC SROCC which is based on both geometry and color features [9]. Re-
sults on sjtu and ICIP20 datasets are shown in Tables 2 and 3,
AlexNet 0.894 ± 0.055 0.874 ± 0.083 respectively with the top-2 results highlighted in bold.
VGG16 0.923 ± 0.031 0.907 ± 0.045
VGG19 0.925 ± 0.035 0.912 ± 0.042 Method PCC SROCC
ResNet18 0.886 ± 0.091 0.856 ± 0.113 Baseline methods (full reference)
ResNet50 0.885 ± 0.086 0.855 ± 0.109 po2pointMSE 0.686 0.801
PSNRpo2pointMSE 0.799 0.844
Table 1. Model comparison on sjtu dataset. The two best
results are highlighted in bold. po2pointHausdorff 0.517 0.686
PSNRpo2pointHausdorff 0.638 0.682
Table 1 shows the mean correlations obtained for each po2planeMSE 0.642 0.717
considered model. The two best results are highlighted in PSNRpo2planeMSE 0.744 0.722
bold. From these results, several observations can be made:
po2planeHausdorff 0.539 0.682
• First of all, the performance differs from a model to PSNRpo2planeHausdorff 0.755 0.825
another with a high gap between them (from 0.875 to po2distMSE (mmd) 0.710 0.603
0.925 in terms of mean PCC and from 0.855 to 0.912
PSNRpo2distMSE (mmd) 0.621 0.603
in terms of mean SROCC). The stability of the re-
sults over the folds differs as well. ResNet-based mod- po2distMSE (msmd) 0.706 0.603
els obtain the higher standard deviations outperformed PSNRpo2distMSE (msmd) 0.642 0.715
by AlexNet, while VGG-based models are more stable
PCQM 0.879 0.888
with the lower standard-deviations.
Our method (no reference)
• ResNet50 achieves the worst performance, lower even
GQI-VGG16 0.923 0.907
than the simpler and shallower model used (i.e.
AlexNet). However, it is competitive to the full ref- GQI-VGG19 0.925 0.912
erence metric PCQM and outperforms all the other full
reference compared metrics on sjtu (see Table 2). Table 2. Comparison with state-of-the-art methods on sjtu
dataset. The two best results are highlighted in bold.
• For VGG-based and ResNet-based models, we can see
that the depth doesn’t highly impact the performance On sjtu (see Table 2), the results obtained by our
since the differences are statistically not significant (p- method with both models (i.e. VGG16 and VGG19)
value > 0.05). surpass all the compared state-of-the-art metrics with a

Authorized licensed use limited to: University Roma Tre AREA SCIENTIFICO TECNOLOGICA. Downloaded on December 11,2023 at 04:20:04 UTC from IEEE Xplore. Restrictions apply.
 gain in terms of mean PCC that varies between 5% and proach). PCQM is outperformed by most of the compared
70%. PCQM achieves the third best performance with metrics, except po2point-based metrics pooled with Haus-
0.879 and 0.888 as mean PCC and SROCC, respectively. dorff (i.e. po2pointHausdorff and PSNRpo2pointHausdorff)
The worst result is obtained by po2pointHausdorff, closely and po2planeHausdorff. Similarly to the results obtained on
followed by po2planeHausdorff. PSNRpo2pointMSE, sjtu, po2point-based and po2plane-based metrics pooled with
PSNRpo2planeHausdorff and po2distMSE obtain the best re- MSE obtain higher correlations than those pooled with PSNR,
sults among the po2point-based, po2plane-based and po2dist- while the po2dist-based metrics pooled with PSNR give better
based metrics, respectively. Whereas po2pointHausdorff, results than those pooled with MSE.
po2planeHausdorff and PSNRpo2distMSE achieve the worst Globally, the correlations reached on ICIP20 are higher
results among the po2point-based, po2plane-based and than those obtained on sjtu, except for PCQM. These results
po2dist-based metrics, respectively. Unlike the po2dist- can be justified by the fact that ICIP20 is composed only
based metrics, po2point-based and po2plane-based metrics of compressed PCs with joint distortion of geometry and at-
pooled with MSE obtain higher correlations than those pooled tributes, while sjtu contains a more wide set of distortions
with PSNR. Globally, the compared metrics, except PCQM, including color noise. We also evaluated the generalization
achieve low correlations since they focus more geometric in- capacity of the proposed method. However, the correlations
formation, failing to catch other distortions. are not as expected.

4. CONCLUSION
Method PCC SROCC
In this paper, we proposed a deep learning-based method that
Baseline methods (full reference)
efficiently predicts the quality of distorted PCs without ref-
po2pointMSE 0.945 0.950 erence. After randomly selecting a set of points from the
PSNRpo2pointMSE 0.880 0.934 PC, we defined a region around each of them by finding the
po2pointHausdorff 0.717 0.690 nearest neighbors. The delimited regions are then character-
ized through three attributes (geometric-distance, mean cur-
PSNRpo2pointHausdorff 0.597 0.763 vature and gray-level) and the obtained values are stacked
po2planeMSE 0.945 0.959 into patches of size 32 × 32 × 3 to predict their quality in-
PSNRpo2planeMSE 0.916 0.953 dexes (i.e. patch quality indexes) through a CNN model. The
global quality index is finally given by aggregating the pre-
po2planeHausdorff 0.753 0.763 dicted patch quality indexes. We compared the performance
PSNRpo2planeHausdorff 0.939 0.970 of five pre-trained CNN models and the best results were com-
po2distMSE (mmd) 0.965 0.963 pared with state-of-the-art methods. The results obtained on
two datasets show the potential of the proposed approach.
PSNRpo2distMSE (mmd) 0.865 0.965
Despite the effectiveness of our method, some points
po2distMSE (msmd) 0.967 0.965 should be deeper analyzed, including the use of more effi-
PSNRpo2distMSE (msmd) 0.902 0.972 cient deep learning models, the impact on the performance of
PCQM 0.796 0.832 the number of selected points as well as the patch size. Other
point selection strategies will also be considered.
Our method (no reference)
GQI-VGG16 0.956 0.966
5. REFERENCES
GQI-VGG19 0.952 0.966
[1] D. Graziosi, O. Nakagami, S. Kuma, A. Zaghetto,
Table 3. Comparison with state-of-the-art methods on ICIP20 T. Suzuki, and A. Tabatabai, “An overview of ongo-
dataset. The two best results are highlighted in bold. ing point cloud compression standardization activities:
video-based (V-PCC) and geometry-based (G-PCC),”
APSIPA Trans. on Signal and Information Process., vol.
On ICIP20 (see Table 3), po2distMSE (msmd) ob- 9, 2020.
tains the best mean PCC (0.967), closely followed by [2] Maurice Quach, Giuseppe Valenzise, and Frederic Du-
po2distMSE (mmd) (0.965). Whereas the two best mean faux, “Learning Convolutional Transforms for Lossy
SROCC is reached by PSNRpo2distMSE (0.972) and Point Cloud Geometry Compression,” in 2019 IEEE
PSNRpo2planeHausdorff (0.970), respectively. Our metric Intl. Conf. on Image Process. (ICIP), Sept. 2019, pp.
achieves the third mean PCC (0.956 for VGG16 and 0.952 4320–4324, ISSN: 1522-4880.
for VGG19) and the fourth mean SROCC (0.966 for both [3] Maurice Quach, Giuseppe Valenzise, and Frederic Du-
models) without accessing the pristine PC (i.e. NR ap- faux, “Improved Deep Point Cloud Geometry Compres-

Authorized licensed use limited to: University Roma Tre AREA SCIENTIFICO TECNOLOGICA. Downloaded on December 11,2023 at 04:20:04 UTC from IEEE Xplore. Restrictions apply.
 sion,” in 2020 IEEE Intl. Workshop on Multimedia Sig- [16] S. Dodge and L. Karam, “Understanding how image
nal Process. (MMSP), Oct. 2020. quality affects deep neural networks,” in 2016 Eighth
[4] Dong Tian et al., “Geometric distortion metrics for point International Conference on Quality of Multimedia Ex-
cloud compression,” in 2017 IEEE Intl. Conf. on Image perience (QoMEX), 2016, pp. 1–6.
Process. (ICIP), Beijing, Sept. 2017, pp. 3460–3464, [17] A. Krizhevsky, “One weird trick for parallelizing convo-
IEEE. lutional neural networks,” CoRR, abs/1404.5997, 2014.
[5] Evangelos Alexiou and Touradj Ebrahimi, “Point Cloud [18] K. Simonyan and A. Zisserman, “Very deep convo-
Quality Assessment Metric Based on Angular Similar- lutional networks for large-scale image recognition,”
ity,” in 2018 IEEE Intl. Conf. on Multimedia and Expo CoRR, abs/1409.1556, 2014.
(ICME), July 2018, pp. 1–6, ISSN: 1945-788X. [19] S. Ren K. He, X. Zhang and J. Sun, “Deep resid-
[6] Gabriel Meynet, Julie Digne, and Guillaume Lavoué, ual learning for image recognition,” arXiv preprint
“PC-MSDM: A quality metric for 3D point clouds,” in arXiv:1512.03385, 2015.
2019 11th Intl. Conf. on Quality of Multimedia Experi- [20] L. Kang, P. Ye, Y. Li, and D. Doermann, “Convolutional
ence (QoMEX), June 2019, pp. 1–3, ISSN: 2472-7814, neural networks for no-reference image quality assess-
2372-7179. ment,” in 2014 IEEE Conference on Computer Vision
and Pattern Recognition, 2014, pp. 1733–1740.
[7] Zhou Wang et al., “Image quality assessment: from er-
[21] Aladine Chetouani, “A blind image quality metric using
ror visibility to structural similarity,” IEEE Tran. on Im-
a selection of relevant patches based on convolutional
age Process., vol. 13, no. 4, pp. 600–612, Apr. 2004.
neural network,” in 2018 26th European Signal Pro-
[8] Guillaume Lavoué, “A Multiscale Metric for 3D Mesh cessing Conference (EUSIPCO), 2018, pp. 1452–1456.
Visual Quality Assessment,” Computer Graphics Fo- [22] Oussama Messai, Aladine Chetouani, F. Hachouf, and
rum, vol. 30, pp. 1427–1437, July 2011. Zianou Seghir, “No-reference stereoscopic image qual-
[9] Gabriel Meynet et al., “PCQM: A Full-Reference Qual- ity predictor using deep features from cyclopean image,”
ity Metric for Colored 3D Point Clouds,” in 2020 01 2021.
12th Intl. Conf. on Quality of Multimedia Experience [23] Ilyass Abouelaziz, Aladine Chetouani, Mohammed
(QoMEX 2020), Athlone, Ireland, May 2020. El Hassouni, LJ Latecki, and Hocine Cherifi, “3D vi-
[10] A. Javaheri et al., “Mahalanobis Based Point to Distri- sual saliency and convolutional neural network for blind
bution Metric for Point Cloud Geometry Quality Eval- mesh quality assessment,” Neural Computing and Ap-
uation,” IEEE Signal Process. Lett., vol. 27, pp. 1350– plications, 2019.
1354, 2020. [24] Scott J. Daly, “Visible differences predictor: an algo-
[11] I. Viola, S. Subramanyam, and P. Cesar, “A Color-Based rithm for the assessment of image fidelity,” in Human
Objective Quality Metric for Point Cloud Contents,” in Vision, Visual Processing, and Digital Display III, Ber-
2020 12th Intl. Conf. on Quality of Multimedia Experi- nice E. Rogowitz, Ed. International Society for Optics
ence (QoMEX), May 2020, pp. 1–6, ISSN: 2472-7814. and Photonics, 1992, vol. 1666, pp. 2 – 15, SPIE.
[12] E. Alexiou and T. Ebrahimi, “Towards a Point Cloud [25] S. Bosse, D. Maniry, K. Müller, T. Wiegand, and
Structural Similarity Metric,” in 2020 IEEE Intl. Conf. W. Samek, “Deep neural networks for no-reference and
on Multimedia Expo Workshops (ICMEW), July 2020, full-reference image quality assessment,” IEEE Trans.
pp. 1–6. on Image Process., vol. 27, no. 1, pp. 206–219, Jan
[13] A. Javaheri et al., “Improving PSNR-based Quality Met- 2018.
rics Performance For Point Cloud Geometry,” in 2020 [26] Q. Yang, Z. Ma, Y. Xu, R. Tang, and J. Sun, “Predict-
IEEE Intl. Conf. on Image Process. (ICIP), Oct. 2020, ing the perceptual quality of point cloud: A 3d-to-2d
pp. 3438–3442, ISSN: 2381-8549. projection-based exploration,” IEEE Trans. on Multi-
[14] Maurice Quach, Aladine Chetouani, Giuseppe Valen- media, 2020.
zise, and Frédéric Dufaux, “A deep perceptual metric [27] Stuart Perry et al., “Quality Evaluation Of Static Point
for 3D point clouds,” in Image Quality and System Per- Clouds Encoded Using MPEG Codecs,” in 2020 IEEE
formance, IS&T International Symposium on Electronic Intl. Conf. on Image Process. (ICIP), Oct. 2020, pp.
Imaging (EI 2021), San Francisco, United States, Jan. 3428–3432, ISSN: 2381-8549.
2021.
[15] Ilyass Abouelaziz, Aladine Chetouani, Mohammed El
Hassouni, Longin Jan Latecki, and Hocine Cherifi, “No-
reference mesh visual quality assessment via ensemble
of convolutional neural networks and compact multi-
linear pooling,” Pattern Recognition, vol. 100, pp.
107174, 2020.

Authorized licensed use limited to: University Roma Tre AREA SCIENTIFICO TECNOLOGICA. Downloaded on December 11,2023 at 04:20:04 UTC from IEEE Xplore. Restrictions apply.