A Simple and Robust Correlation Filtering method for text-based person search

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).

Getting Started

Requirements

PyTorch 1.4 or higher
transformers (install with pip install transformers)
numpy, torchvision

Dataset Preparation

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

Building BERT

mkdir bert_weight

Downland the weight and config, put them into SRCF/bert_weight

Training and Testing

bash run/train.bash

Evaluation

bash run/test.bash

Results on CUHK-PEDES

CUHK-PEDES	performance
`Top-1`	64.88
`Top-5`	83.02
`Top-10`	88.56

Citation

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}
}

References

SSAN

Suo-Wei/SRCF