Object Detection using Convolution Neural Network

Overfeat
 Overfeat
Idea 4 Sliding window (spatial output) + Pretrained Model + Image Pyramid + FC as Convnet

1. Train Localize multi class classification - To understand generalize property of

model Bbox

2. Use it as pretrained model for object - Multi class detection

detection task

3. Sliding window - To detect multiple instance of object Overfeat

Paper
4. Image Pyramid - For detecting varying size object

5. FC as convnet - To overcome CNN constrain

6. NMS - For final prediction

 Overfeat
Idea 4 Sliding window (spatial output ) + Pretrained Model + Image Pyramid + FC as convnet Testing

FC as convnet

Regression

MODEL
F

VGG 16
C C
1 2

Sliding window Feature Extraction

+ Classification
Image Pyramid
Pre Trained Localize Model
Overfeat - Experiment

~ winner of the localization task of the ImageNet Large

Scale Visual Recognition Challenge 2013 (ILSVRC2013)

Paper link: https://arxiv.org/abs/1312.6229

 Overfeat - Experiment - Classification and Localization - Training

2012

• ImageNet 2012 dataset

• 1000 classes

• Trained classification and localization

task on modified AlexNET.
Vehicle Dog Craft • ILSVRC 2013 1st Winner for
classification and localization task

• 3rd Rank for Detection task

1000 classes
 Overfeat - Experiment - Classification and Localization - Training

Classification and Localization

Localize
Bounding Box
ImageNet 2012
Dataset – 1000 CLASS Regression

MODEL
F
C C
1 2

Feature Extraction
Image + Bounding Box Classification

Modified AlexNET Classify Class k

• ILSVRC 2013 1st Winner for classification and localization task

 Overfeat - Experiment - Object Detection - Training

• Training without Background

class leads to a lot of False
Positive prediction.

• To avoid FP, one additional

class is used.

• Training data for background

245 x 245 is taken randomly where no
object is present.
Classification and Localization
• Training of Classification and
Regression Localization is done on 20 + 1
Class.
MODEL

Training
C + 1 Class
• Base dimension for training is
+ 1 for Background Classification 245 x 245
 Overfeat - Experiment - Object Detection - Inference

• Pretrained Classification and Localization Network

trained on C + 1 Class

• Spatial output of Image pyramid with base dimension

of prediction is 245 x 245

• 6 Scale Image pyramid with 1:2 factor of Resolution

• FC as Convent

• Resolution / Subsampling ratio / Effective Strides = 36

 Overfeat - Experiment - Resolution

Resolution / Subsampling Ratio / Effective Stride = 36 Resolution / Subsampling Ratio / Effective Stride = 18

389 x 461 317 x 386 317 x 386

Spatial Output 3x5 6 x 10

For 1 class only
5x7
 Overfeat - Experiment - Image Pyramid - Spatial Output
6 Scale Image Pyramid with 1:2 factor of Resolution

281 x 317 317 x 386 389 x 461 425 x 497 464 x 569

2 x 3 x C+1

3 x 5 x C+1 5 x 7 x C+1
6 x 7 x C+1
7 x 10 x C+1

1 x 1 x C+1 Resolution / Effective Stride = 36

245 X 245
Credit: https://www.youtube.com/watch?v=9I6nzfx_kpE&list=PL1GQaVhO4f_jLxOokW7CS5kY_J1t1T17S&ab_channel=Cogneethi
 Overfeat - Experiment - Object Detection - Inference
Spatial Output

2x3 + 3x5 + 5x7 + 6x7 + 7 x 10 = 169 x C+1 Resolution / Effective Stride = 36

NMS

245 X 245 245 X 245

 Overfeat - Experiment - Object Detection - Inference
1 x 1 x 4 x c+1
245 X 245
2 x 3 x 4 x c+1
Localize
Bounding Box 3 x 5 x 4 x c+1
281 x 317
Regression 5 x 7 x 4 x c+1
Classification and Localization
317 x 386 6 x 7 x 4 x c+1
Query 7 x 10 x 4 x c+1

MODEL
389 x 461
245 X 245

Modified Alex NET 1 x 1 x c+1

425 x 497
2 x 3 x c+1
Classification
3 x 5 x c+1
Classify
Class c+1 5 x 7 x c+1
464 x 569
6 x 7 x c+1
7 x 10 x c+1
Credit: https://www.youtube.com/watch?v=9I6nzfx_kpE&list=PL1GQaVhO4f_jLxOokW7CS5kY_J1t1T17S&ab_channel=Cogneethi
Query Input Resolution = 36

Classification
Confidence Box Resolution = 12

Model

Regression
Bounding Box

Credit: https://arxiv.org/abs/1312.6229
 Overfeat - Experiment - Object Detection - Inference

NMS
Spatial Outputs

Final prediction

Credit: https://arxiv.org/abs/1312.6229
Overfeat - Experiment - Object Detection - Research Paper Result

• 3rd Rank for Detection task

Paper link: https://arxiv.org/abs/1312.6229

Overfeat - Object Detection - Drawbacks
Background Class

Model Background Class

 Overfeat - Object Detection - Drawbacks

• Each inference takes 2 seconds

• Computationally Inefficient and also expensive (2013)

• A lot of background region is getting unnecessary processed because of Sliding

window or dense Sampling approach

• Sliding window approach creates a lot of FP , therefore less MAP

• Can we have some way to find to only those regions where background is not present?

• Accurate predictions and increase MAP ?

• RPNN END 
Credit :https://www.pyimagesearch.com/2020/06/29/opencv-selective-search-for-object-detection/