The goal of this work is to improve the performance of global text to image person retrieval under weakly supervised settings. To achieve this, we utilize the complete CLIP model as our feature extraction backbone. In addition, we propose an Image-Centered Pseudo Label Generation method to address the issues of uncertain cross-modal pseudo identity labels.

Download the CUHK-PEDES dataset from here, ICFG-PEDES dataset from here and RSTPReid dataset form here

Organize them in your dataset root dir folder as follows:

|-- your dataset root dir/
|   |-- <CUHK-PEDES>/
|       |-- imgs
|            |-- cam_a
|            |-- cam_b
|            |-- ...
|       |-- reid_raw.json
|
|   |-- <ICFG-PEDES>/
|       |-- imgs
|            |-- test
|            |-- train 
|       |-- ICFG_PEDES.json
|
|   |-- <RSTPReid>/
|       |-- imgs
|       |-- data_captions.json

python train.py \
--name icpg \
--img_aug \
--batch_size 64 \
--MLM \
--loss_names 'chm+cdm+itc' \
--dataset_name 'CUHK-PEDES' \
--num_epoch 60

python test.py --config_file 'path/to/model_dir/configs.yaml'