The Constants I chose are as follows:
These parameters were chosen based on the tradeoff of performance and runtime. I found that after initially using a negative features number of 50,000 greatly increased the runtime and was not worth it. My implementation performed well on most images in finding faces, although there were many instances in which it classified non-faces as faces, or false positives. My average precision was 0.829.
Face template HoG visualization for the starter code. This is completely random, but it should actually look like a face once you train a reasonable classifier.
Precision Recall curve for the starter code.