Towards precise intra-camera supervised person re-identification

The official implementation for Precise-ICS (Towards precise intra-camera supervised person re-identification), accepted to WACV 2021.

Preparation

Requirements: Pytorch>=1.1.0 and python>=3.6

1. install pytorch
2. Download re-ID dataset
   • Market-1501
   • DukeMTMC-reID
   • MSMT17
3. Put the data under the dataset directory. Training, query and test sub-folder should named as bounding_box_train, query, bounding_box_test, respectively.

Training and test model for Precise ICS

# example: train Precise-ICS model on Market-1501
CUDA_VISIBLE_DEVICES=1 python train.py --log_path train_full_model.txt --save_model_interval 50 --dataset 'market1501' --market_path '/path/to/dataset/Market1501/' --associate_rate 1.5

Note:

If necessary, you can manually set other hyper-parameters in the training script, such as T (for temperature) and miu (for updating rate) in memory bank learning.

Associate rate is set in the range of [0.5, 1.5] for better performance. In experiments we set 1.5 for Market-1501 and 1.0 for both DuekMTMC-ReID and MSMT17.