Face Recognition

by Eli Bosworth (eboswort)

---

The first step for my face recognition was to build a better feature that a raw image crop for representing face-iness. I used a SIFT descriptor, I experimented with using more than one SIFT descriptor but I saw no real improvement, and a increase in running time, so I stuck with one. Using this SIFT descriptor and a linear classifier with random negatives, this was my result, .325 Average Precision:

The next thing I experimented with was a non-linear classifier. I tried specifying different sigma values, but I found I got my best results by not specifying the sigma and allowing the SVM to choose it's own sigma, which gave me .450 precision:

I then added hard negatives rather then just using random negatives. I found I got better results by using hard negatives in addition to random negatives, rather than replacing the random ones, but my best result was only slightly better than my best result with random negatives. Maybe I just got some really good random negatives? This was my best result, .462 average precision: