We provide the code for reproducing results of our ECCV 2022 paper A Simple and Robust Correlation Filtering method for text-based person search. Compared with the original paper, we obtain better performance due to some modifications. Following our global response map, we also add the same mutual-exclusion-loss to separate body part response map. Meanwhile, we merge the global filter and dictionary filter module. The adjusted method achieves a new state-of-the-art performance and it improves to 64.88 on Top-1 without Re-Rank (CUHK-PEDES).
- PyTorch 1.4 or higher
- transformers (install with
pip install transformers) - numpy, torchvision
Organize them in dataset folder as follows:
|-- dataset/
| |-- <CUHK-PEDES>/
| |-- imgs
|-- cam_a
|-- cam_b
|-- ...
| |-- reid_raw.json
Download the CUHK-PEDES dataset from [here](https://github.com/ShuangLI59/Person-Search-with-Natural-Language-Description) and then run the `process_CUHK_data.py` as follow:
cd SRCF
python ./dataset/process_CUHK_data.py
mkdir bert_weight
Downland the weight and config, put them into SRCF/bert_weight
bash run/train.bash
bash run/test.bash
| CUHK-PEDES | performance |
|---|---|
Top-1 |
64.88 |
Top-5 |
83.02 |
Top-10 |
88.56 |
If this work is helpful for your research, please cite our work:
@InProceedings{Suo_ECCV_A,
author = {Suo, Wei and Sun, MengYang and Niu, Kai, et.al},
title = {A Simple and Robust Correlation Filtering method for text-based person search},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {August},
year = {2022}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent