matchbox

introduction

This tool tries to detect whether the test images are from the same person as the training images.

steps

1. extract feature representations from training faces
2. calculate mean (or median)
3. extract feature representations from a test face
4. compare and calculate a distance estimate
5. images with a distance within the threshold are considered to be of the same person

usage

Tested in docker container (bamos/openface). example:

python main.py --verbose  --targetDir /path/to/rawimg/Yifei_Liu/ /path/to/rawimg/Wei_Zhao/006-vi.jpg /path/to/rawimg/Wei_Zhao/007-vi.jpg | tee zhaowei-liuyifei.log

TODO:

improve distance estimation: instead of comparing test image with mean representaions, try to compare the test image with each training image, then check how many (80%-90%) of them pass a certain threshold.