GitXplorerGitXplorer
l

Selective-Search

public
0 stars
0 forks
0 issues

Commits

List of commits on branch master.
Unverified
7265df75b40fbfc114293862940ca99aa5fbfca1

[Add] comment and format code.

lliuliuOD committed 5 years ago
Unverified
23908082bc3ac243e84bde01200c748bdf0c930f

[Modify] description of color similarity in README.md.

lliuliuOD committed 5 years ago
Unverified
2036893d2c55300382a3cb0bc2bd24b73ec626dc

[Format] more_SS.py.

lliuliuOD committed 5 years ago
Unverified
ebf45018084920e8a53542785ed3d589e5355ba4

[Fix] README.md img path error and case sensitive.

lliuliuOD committed 5 years ago
Unverified
57e93d571d5f094ed0cf902ba9b4fd53b95d8c20

[Modify] directory structure.

lliuliuOD committed 5 years ago
Unverified
7105d32ae3a8ae28d388e149209b054946c8df42

Fix readme.md letter wrong.

CChialiuliu committed 6 years ago

README

The README file for this repository.

Selective-Search

  • This approach is used to get region proposals in Fast-RCNN.
  • for more detail

Setup

  • Windows 10
  • Python 3.6

Result

Detail

1. Selective Search uses graph-based image segmentation to get initial regions, first.
In my practice, I just used the function of graph-based image segmentation in openCV.

2. Calculate similarities between each neighboring region.

Following are 4 similarity approaches in this paper.

A. color similarity:

  In order to measure the similarity in color, this approach can measure color distribution  
  of each region.  
  Calculate color histogram of each region, than sum up the minimize interval between both  
  neighboring regions.  

  We can easily get the color similarity after merge two region by using following algorithm.

B. texture similarity:

  Similar objects have more similar texture features, so we measure the texture similarities  
  between each neighboring region to merge complete object.  
  In this paper, they use SIFT to get image textures.  
  In my practice, I used LBP instead.  

C. size similarity:

  To avoid all regions be merged by a big region.  

D. fill similarity:

  Merge the regions which has the least hole after merging.  

Following are two bounding boxes with same widths and heights, show what is the least hole.

3. Merge the highest similarity of two regions, iteratively.