Just another sliding window algorithm to detect faces of variable dimensions using HOG descriptors as input to a SVM model. The dataset can be found here (Caltech 10, 000 Web Faces). Using data augmentation, tuning some paramters such as the hog cell size, the downscale factor for the sliding window and the detection threshold I managed to achieve a mAP of 0.833.
