Triplet Margin Loss for Person Re-identification

This Project is for Person Re-identification using Triplet Loss based on PyTorch

• Triplet Margin Loss
• load weights form darknet weights file
• save weighes file with darknet format
• find best threshold for test set
• more network structure
• more trick using in ReID
• faster in multi-GPU
• load and save caffemode

Training and validation

1. creat triplet list file and put it in data/
2. set image_root=your/images/path

python train_val.py --gpus 0,1,2,3

Time Cost

time cost in 4 TITAN X

batch size	cost(ms) / 1 TripletSample
64	1.31
128	0.96
256	0.35
512	0.20
1024	0.18

Reference

• FaceNet: A Unified Embedding for Face Recognition and Clustering