A PyTorch based strong baseline for person search

This repository develops a baseline model with high performance, and implements the widely used baseline OIM [1], and NAE [4].

• Separating detection and re-ID head on the top of model
• PK sampling for training re-ID head
• Warm-up training

About this repository

• It is pure PyTorch code, which requires the PyTorch version >= 1.1.0
• It supports multi-image batch training.
• End-to-end training and evaluation. Both PRW and CUHK-SYSU are supported.
• Standard protocol (including PRW-mini in [3]) used by most research papers
• Highly extensible (easy to add models, datasets, training methods, etc.)
• Visualization tools (proposals, losses in training)
• High performance baseline.

Preparation

• Clone repository and build the environment

  git clone https://github.com/DeepAlchemist/deep-person-search.git && cd deep-person-search
  conda env create -f env.yml

• Download PRW and CUHK-SYSU (also named SSM) to path_to_your_data.

• In the config file ./lib/cfg/config.py, change --data_root to path_to_your_data, --ckpt_dir to path_you_want_to_save_the_checkpoints.

• Download ImageNet pre-trained ResNet models from GoogleDrive to deep-person-search/cache/pretrained_model/

Training & Evaluation

• Training

CUDA_VISIBLE_DEVICES=0 python main.py \
    --benchmark ssm --batch_size 5 \
    --backbone bsl --in_level C5 --cls_type oim \
    --lr 0.003 --warmup_epochs 1 --max_epoch 7 \
    --suffix "" 

-dis enable display (visualization), then tensorboard --bind_all --logdir your_log_dir, which shows the loss curves and the input image with proposals.

• Evaluation

CUDA_VISIBLE_DEVICES=0 python main.py --is_test \
    --benchmark ssm --eval_batch_size 5 \
    --backbone bsl --in_level C5 --cls_type oim \
    --load_ckpt "absolute_path_to_your_checkpoint" \

• Comparison

The download link of the trained models are available in the table. Note that all the models are trained with image size of 600x1000, the larger image size, e.g., 900x1500, would yield better performance.

TODO

• DistributedDataParallel
• Training with larger image size, i.e., 900x1500
• Supporting more SOTA methods
• Visualizing ranking list in test
• A technological report for this repository

[1] Joint Detection and Identification Feature Learning for Person Search. In CVPR 2017.
[2] Person Re-Identification in the Wild. In CVPR 2017.
[3] Query-guided End-to-End Person Search. In CVPR 2019.
[4] Norm-Aware Embedding for Efficient Person Search. In CVPR 2020.

Citation

@article{yang2021bottom,
  title={Bottom-up foreground-aware feature fusion for practical person search},
  author={Yang, Wenjie and Huang, Houjing and Chen, Xiaotang and Huang, Kaiqi},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  volume={32},
  number={1},
  pages={262--274},
  year={2021},
  publisher={IEEE}
}