In 2005 at the International Conference on Computer Vision, Li Fei-Fei , Rob Fergus, and Antonio Torralba collaborated on an award-winning workshop about object recognition. The materials from that workshop are available from Torralba's web page, and include three Matlab demos you can try out.
Bag of Words
You will need to unzip the files first and store them somewhere accessible to Matlab. Next you will need to modify the path and also locations within several files so that they point to the correct locations. Specifically, edit the values of RUN_DIR, IMAGE_DIR, and CODEBOOK_DIR within config_file_1.m and config_file_2.m. On Windows you should also create a temporary directory and edit lines 90-93 of gg_lola_km_binary.m to point to it. Also, add the command scale = 0; at line 18 of SIFT.m.
I got stuck trying to run this on Windows -- it tries to run a binary file that seems to be designed for Unix. On reading the documentation more carefully, I see that this restriction was actually noted in the web page. In theory it should be runnable in unix, but I have not tried.
Part Structured Models
Similar to Bag of Words, modify the values of RUN_DIR, IMAGE_DIR, and CODEBOOK_DIR within config_file_1.m and config_file_2.m.
There are a few small changes that seem necessary before the file will run without errors. In do_test_parts_structur make the following changes:
- On line 373 add try
- On line 424 add catch; end
- Change line 427 from pause to drawnow
Boosting Demo
The instructions for the first part of this one are pretty straightforward. Unzip everything, edit initpath.m and parameters.m, then run demoGentleBoost.
When you are ready for the second part, you will need to add all the subfolders of the LabelMe package to the Matlab path. I had to make a few changes to the supplied code in order to get it to run without errors.
- Change line 70 of runDetector to the following: plotBoundingBox(newimg,boundingBox(sb,:), fix(boxScores(sb)/40)+1)
- Change line 23 of plotBoundingBox to the following: plot([bb(i,1) bb(i,3) bb(i,3) bb(i,1) bb(i,1)], [bb(i,2) bb(i,2) bb(i,4) bb(i,4) bb(i,2)], colors(mod(i,6)+1), 'linewidth', max(1,score(i)-2))
With these changes the detector code seems to run, although it appears to me that its grading function is incorrect. (It seems to be saying that some bounding boxes are correct even when they don't look right.)