Revision history [back]

I spent some time reading the FPDW and VeryFast papers. The conclusion is that they are mostly performance optimizations and small tweaks, while the fundamental detector is still HOG. Also, the classifier (SVM, NN, decision forest) does not make a huge difference. Therefore, I decided to go with HOG+SVM because it's fast and simple, and I found a good implementation from http://dlib.net. Example: http://blog.dlib.net/2014/02/dlib-186-released-make-your-own-object.html Dlib uses an improved version of HOG, by Felzenszwalb et al. Also, performance _should_ be better than OpenCV.