SelectiveSearchForWSOD

You can use this project to generate the proposals (pkl) file for your own dataset.

It works well in PCL and WETECTRON.

Overview

0、prepare your own dataset, this project only need pictures and the split file "train.txt/test.txt/val.txt" used in VOC format.

1、Use this Project to generate proposals by Matlab

note: I made some modifications to it.

2、merge the regions into a pkl file by a simple python program

Get Started

1、Download this Project And use its demo to make sure it works

2、Replace it's start.m with start.m, I make some modifications to do following things

1、In pkl file, proposals list of an image needs a index to point out with picture it belongs to.
2、The boxes are in [ymin, xmin, ymax, xmax] format, I transfer them into [xmin,ymin,xmax,ymax]

3、make a new folder named image in SelectiveSearchCodeIJCV and put the images of your dataset into it

4、put the split file (ie. test.txt;val.txt etc.) into SelectiveSearchCodeIJCV

Now, Your project should be like

--SelectiveSearch-RCNN
  --SelectiveSearchCodeIJCV
    --image
      the images of your dataset 
    train.txt(if needed)
    test.txt(if needed)
    val.txt(if needed)
    trainval.txt(if needed)
    ...and others

5、change this line to specify the subset to be generate and run start.m.

Note: remember to save the out_box.mat & out_index.mat to another place everytime you run and make sure you know which split they below to. Otherwise they will be overwrite

for example:
idx_file = './SelectiveSearchCodeIJCV/train.txt'; for training set
idx_file = './SelectiveSearchCodeIJCV/test.txt'; for test set
idx_file = './SelectiveSearchCodeIJCV/val.txt'; for validation set

6、merge two .mat into a pkl file with mergeRegions.py.

out_box.mat and out_index.mat need to be in the same folder with mergeRegions.py

Note: Now you have the pkl for one subset in the whole dataset, you need to repeat from step 5 for every subset you need