R-CNN 정리(2)

논문 원제목)

Rich feature hierarchies for accurate object detection and semantic segmentation

 

논문 개인 공부 기록 남기기 포스팅입니다.

 

전반적 중요 내용 요약이므로 논문 전체의 디테일 요소까지 

 

다루지 않을 수 있음을 알려드립니다.

 

 

 

[Object detection with R-CNN]

Object Detection on this paper is consisted of 3 Modules.

 

Three Modules)

1. Category-Independent region proposals are generated.(Not preprocessed)

2. Large CNN->Output : Feature Vector with fixed length based on feature map for each region.

3. Class-specific linear SVMs (Classification).

 

<Module Design>

Region Proposals)

Used "Selective Search Algorithm" to enable controlled comparison with prior detection work.

 

Feature Extraction)

4096-dimensional feature vector from each region proposal will be extracted. (Based on Caffe Model)

 

Feature extraction is based on forward propagating 

227 x 227 sized RGB image through 5 conv layers and 2 FC layers.

(Model Architecture is based on this paper.  A. Krizhevsky, I. Sutskever, and G. Hinton. ImageNet classification
with deep convolutional neural networks. In NIPS, 2012.)

 

 

Each region proposal has different sizes respectively.

 

Since CNN architecture requires fixed size of input data image, 

we have to warp(resize) each of proposed region into 227 x 227 pixel size.

 

Image Reference URL(bskyvision님 감사합니다) : bskyvision.com/694

Example of Region Proposal & Warping operation

 

 

 

 

Q : How does warp operation processed?

 

A : Warp all pixels in a tight bounding box around it to the required size.

     Also, during this process some bit of information could be loss.

*Note* Prior to warping, dilation operation is applied to tight bounding box. ->So that at the warped size, there would be exactly p pixels.

 

 

 

 

 

[Test-time detection]

At test time, region proposal (Extract 2000 images) has processed based on "Selective Search Algorithm".

That "Selective Search Algorithm" has been performed with fast mode.

 

 

Q : What is fast mode in Selective Search Algorithm? A : It's one of diverse modes that Selective Search Algorithm(SSA) adopted.        First, In the fast mode, color space is limited to HSV and Lab.        Lab is abbreviated word which has three meaning.    1. L => Lightness    2. a => 2 conditions exist.              if a<0 : then it's biased Green color.              else   :  then it's biased Red Color.    3. b==> 2 conditons exist.              if a<0 : then it's biased Blue color.              else   :  then it's biased Yellow Color.   Second, Similarity Measures : CTSF   Third , starting regions are 50,100   Reference URL : pypi.org/project/selective-search/

    

And, we scored each extracted feature vector using SVM.(Classifier)

Also, applied "Greedy non-maximum suppression"

 

=> Remove regions which are overlapped IoU with a higher scoring

     selected region larger than a learned threshold.

(overlapping이 많은 region들은 IoU 수치가 높을 것이며

해당 IoU 수치가 일정 Threshold 값 이상이면 해당 region proposal을

삭제하여 최적의 Bounding Box만 남기는 작업이다)

 

Image Reference URL : ballentain.tistory.com/12

Greedy non-maximum suppression

 

 

 

Run-time Anlaysis)

There are 2 properties make detection efficient.

 

Q) What are those 2 properties?

A)  1. All CNN parameters are shared across all categories.

     2. Compared to previous approach, feature vectors computed by CNN are low-dimensional