This repository provides the official PyTorch implementation of Bi-IRRA.

• Release of code and multilingual annotation JSON files
• Release of pretraining checkpoints

All experiments are conducted on 4 Nvidia A40 (48GB) GPUs with CUDA 11.7.

Install the required packages using:

pip install -r requirements.txt

1. Datasets

   • Download CUHK-PEDES from here
   • Download ICFG-PEDES from here
   • Download RSTPReid from here
   • Download UFineBench from here

2. Annotations

   • Download the multilingual annotation JSON files from here

3. Pretrained Models

   • Download the xlm-roberta-base checkpoint from here
   • Download the CCLM-X2VLM checkpoint from here
   • We also provide the checkpoint pretrained on the LUPerson dataset, which can be used as a replacement for the CCLM-X2VLM checkpoint to achieve better retrieval performance.

Edit config/config.yaml (for CUHK-PEDES, ICFG-PEDES, RSTPReid) or config/config_UFine.yaml (for UFineBench) to set the paths for annotation files, image directories, tokenizer, and checkpoints.

A recommended directory structure:

|- CUHK-PEDES
   |- imgs
   |- annos
      |- train_reid_en.json
      |- train_reid_ch.json
      |- test_reid_en.json
      |- test_reid_ch.json

Example configuration in config.yaml:

anno_dir: /path/to/CUHK-PEDES/annos
image_dir: /path/to/CUHK-PEDES/imgs
text_encoder: /path/to/xlm-roberta-base
resume: /path/to/CCLM-X2VLM

Start training with:

CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --rdzv_id=3 --rdzv_backend=c10d --rdzv_endpoint=localhost:0 --nnodes=1 --nproc_per_node=4 \
main.py

or simply run:

bash shell/train.sh

This repository is partially based on TBPS-CLIP and X2-VLM. We thank the authors for their contributions.

This project is licensed under the MIT License.