Adaptive-Exploration-for-Unsupervised-Person-Re-Identification

Prerequisites

• Python 3.6
• Pytorch 1.0

Datasets

1. Create folder to save data mkdir data.
2. Download the datasets (Market-1501, DukeMTMC-reID and MSMT17). If you want dataset from Baidu Yun, please refer to ECN (Thanks to Zhun Zhong).
3. Unzip them and put the unzipped file under data/.
4. The data structure would look like:

data/
    market/
          bounding_box_train/
          bounding_box_test/
          bounding_box_train_camstyle/
          query/
    duke/
          bounding_box_train/
          bounding_box_test/
          bounding_box_train_camstyle/
          query/
    msmt17/
          bounding_box_train/
          bounding_box_test/
          bounding_box_train_camstyle/
          query/

Train

run bash train.sh.

Test

we also provide the pretrained model for testing.

run bash test.sh.

Results (paper)

1. Market1501(market) and DukeMTMC-reID(duke)

Method & data	Map	rank-1	rank-5	rank10
duke to market	58.0%	81.6%	91.9%	94.6%
market only	54.0%	77.5%	89.8%	93.4%
market to duke	46.7%	67.9%	79.2%	83.6%
duke only	39.0%	63.2%	75.4%	79.4%

2. MSMT17(msmt17)

Method & data	Map	rank-1	rank-5	rank10
market to msmt17	9.2%	25.5%	37.3%	42.6%
duke to msmt17	11.7%	32.3%	44.4%	50.1%
msmt17 only	8.5%	26.6%	37.0%	41.7%

Citation

If you find the code useful, considering citing our work:

@article{journals/tomccap/DingFXY20,
  author    = {Yuhang Ding and Hehe Fan and Mingliang Xu and Yi Yang},
  title     = {Adaptive Exploration for Unsupervised Person Re-Identification},
  journal   = {{TOMM}},
  volume    = {16},
  number    = {1},
  pages     = {3:1--3:19},
  year      = {2020},
  doi       = {10.1145/3369393},
}

Related Repos

https://github.com/zhunzhong07/ECN

https://github.com/Cysu/open-reid

https://github.com/layumi/Person_reID_baseline_pytorch