Channel-Aware Cross-Fused Transformer-style Networks (C2T-Net)

Doanh C. Bui, Thinh V. Le and Hung Ba Ngo

1. Prepare the dataset from UPAR challenge

Follow the instruction of data download of organizer website.

The data should be arranged as below tree directory:

data
├── phase1
│   ├── annotations
│   ├── Market1501
│   ├── market_1501.zip
│   ├── PA100k
│   └── PETA
├── phase2
│   ├── annotations
│   ├── MEVID
│   └── submission_templates_test

2. Prepare docker image

Download docker image here.

Run the below command to load the docker image:

sudo docker load < upar_hdt.tar

Go into the data folder, run below command to create a container

sudo docker run -d --shm-size 8G --gpus="all" -it --name upar_hdt --mount type=bind,source="$(pwd)",target=/home/data upar_hdt:v0

Run the container

sudo docker exec -ti upar_hdt /bin/bash

Then, follow the step 3 for reproducing the results, and step 4 for training.

3. Inference for testing dataset in phase 2:

Download our best checkpoint here (best_model.pth). Place it under checkpoints folder (we already put it in the docker image).

Run the below file for inference:

CUDA_VISIBLE_DEVICES=0 python infer_upar_test_phase.py

The results are written in predictions.csv file. This file is valid for submission in the codalearn portal.

3. Training model

Run the below command for training:

CUDA_VISIBLE_DEVICES=0 bash run.sh

The checkpoints and logs would be saved at exp_results/upar/

If this repository proves beneficial for your projects, we kindly request acknowledgment through proper citation:

@InProceedings{Bui_2024_WACV,
    author    = {Bui, Doanh C. and Le, Thinh V. and Ngo, Ba Hung},
    title     = {C2T-Net: Channel-Aware Cross-Fused Transformer-Style Networks for Pedestrian Attribute Recognition},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
    month     = {January},
    year      = {2024},
    pages     = {351-358}
}