C++ framework for Face Matching v2.0

This package contains codes for extracting features from normalized images and computing matching scores from lists of gallery and probe images.

Requirement

1. CUDA 8.0
2. CUDNN 6.0
3. Tensorflow (included in the 3rdparty folder)
4. Mxnet 1.0 (included in the 3rdparty folder)
5. OpenCV3.4.0

Installation/Compilation

cd sample_codes/FaceMatcher_main
make clean; make

Usage

• Step 1: Matching gallery and probe, run the following: ./run_matching.sh <gallery_list_file> <probe_list_file> <output_score_file>

cd bin/facematcher
./run_matching.sh ../../sample_images/small_test/list_gallery.txt ../../sample_images/small_test/list_probe.txt ../../results/score_mat.txt

• Step 2: Compute accuracy from score matrix, run the following: ./run_get_accuracy_all_threshold.sh <gallery_list_file> <probe_list_file> <score_file> <output_accuracy_file>

./run_get_accuracy_all_threshold.sh ../../sample_images/small_test/list_gallery.txt ../../sample_images/small_test/list_probe.txt ../../results/score_mat.txt ../../results/acc_all.txt

NOTE: These examples are for small testing set, all the probe images are from hard error cases we received previously