Awesome

Rank Re-identification

Introduction

• Implement Siamese network proposed by Zhedong in Keras
• Using RankNet to regress ranking probability

Enviroment

python2.7, tensorflow 1.4 and keras 2.1.5

Model

Siamese Network

What's new

• Refer to Zhedong's pytorch baseline, we write a memory-friendly evaluation script, when you generate ranking id tables, for example named with 'market_rank_pid.txt', you can use python3 eval_on_result.py --target_dataset_path /home/cwh/coding/dataset/market --pid_path market_rank_pid.txt --result_path market_result.log to compute map and rank1, rank5, rank10. （test on python3, pytorch 0.4.0）

RankNet

• Base Network：ResNet50
• Input：Given a ranking list, choose a probe image A, two gallery image B, C
• Output：Compute the probability that rank AB > rank AC

Hardware

• NVIDIA TITANX 11G
• Memory: >=16G

Data prepare

Download

• CUHK01
• VIPeR
• Market-1501
• GRID

Preprocess

Take Market-1501 as an example:

• download
• rename training directory to 'train', rename probe directory to 'probe', renmae gallery directory to 'test'

Your data directory will look like this:

Market-1501
├── probe
│   ├── 0003_c1s6_015971_00.jpg
│   ├── 0003_c3s3_064744_00.jpg
│   ├── 0003_c4s6_015641_00.jpg
│   ├── 0003_c5s3_065187_00.jpg
│   └── 0003_c6s3_088392_00.jpg
├── test
│   ├── 0003_c1s6_015971_02.jpg
│   ├── 0003_c1s6_015996_02.jpg
│   ├── 0003_c4s6_015716_03.jpg
│   ├── 0003_c5s3_065187_01.jpg
│   ├── 0003_c6s3_088392_04.jpg
│   └── 0003_c6s3_088442_04.jpg
└── train
    ├── 0002_c1s1_000451_03.jpg
    ├── 0002_c1s1_000551_01.jpg
    ├── 0002_c1s1_000776_01.jpg
    ├── 0002_c1s1_000801_01.jpg
    ├── 0002_c1s1_069056_02.jpg
    └── 0002_c6s1_073451_02.jpg

• replace all paths containing '/home/cwh/coding' to your data path in these file:
  • baseline/train.py
  • pretrain/eval.py
  • pretrain/pair_train.py
  • transfer/simple_rank_transfer.py

Execution

• Train a Resnet-based softmax classifer: specify source dataset and run baseline/train.py
• Train a siamese network: specify source dataset and run pretrain/pair_train.py
• Train a RankNet: refer to TFusion

Detail Instruction

• baseline：ResNet52 base network

  • evaluate.py
    • extract_feature: compute ranking result by base network and evaluate rank accuracy + mAP
    • similarity_matrix: Cosine similarity of CNN features(accelerated by GPU)
    • On test set, compute ranking table with test_predict
    • On training set，compute ranking table with train_predict
    • Compute rank accuracy and mAP with map_rank_quick_eval on Market1501(can be extended to DukeMTMC4ReID)
    • Compute rank accuracy with grid_result_eval on GRID
  • train.py
    • Use source dataset to pretrain ResNet52 base network

• pair: pretrain siamese network

  • pair_train.py：pretrain with two input images
    • pair_generator: data generator, selecting positive and negative samples according to person id
    • pair_model: build a Keras based Siamese network
  • eval：evaluate on Siamese Network and ranknet
    • load corresponding model
    • call function in baseline/evaluate.py for test

• transfer: incremental training with ranking table

  • simple_rank_transfer.py: learning to rank with three input images
    • triplet_generator_by_rank_list：image generator
    • rank_transfer_model：three input image, one ranking loss

Reference

• RankNet

Citation

Please cite this paper in your publications if it helps your research:

@article{
  title={Unsupervised Cross-dataset Person Re-identification by Transfer Learning of Spatial-Temporal Patterns},
  author={Jianming, Lv and Weihang, Chen and Qing, Li and Can, Yang},
  journal={CVPR},
  year={2018}
}