|
I just thought it would be a good idea to have a post to list the image features (or code to generate them) that are publicly available for the common image benchmark datasets (MSRC, PASCAL VOC, imagenet,...). I will start with the Fully Connected CRFs paper published at NIPS last year. They have both Textonboost features and code for the MSRC-21 and PASCAL VOC 2010 datasets : http://graphics.stanford.edu/projects/densecrf/ |
|
Hi Aurelien. What language where you thinking about? Most of the Pascal things are unfortunately written in Matlab. I think texton boost will be more helpful for MSRC than for Pascal but you can give it a try. The most commonly used features for Pascal are (color) sift bag of word representations with spatial pyramids as well as histograms of oriented gradients. Both are usually used together with chi2 or intersection kernels. The Pascal website lists some useful software. For SIFT I usually use vl_feat which is mainly matlab but also has Python bindings. Doing colorsift (rg or opponent) is pretty easy given the gray scale sift from vl_feat, but there is also a closed source binary for doing that. I'm not sure if there is a standard implementation of hog, but I know there is one for python in scikits-image. Other interesting descriptors are local binary patterns and brief, but both haven't really been used much on Pascal AFAIK. If you are a Matlab user, the code from Parreira and Sminchisescu is pretty interesting. The compute phog and pbow. |
