This is an official implementation of the AAAI-2021 paper "KGDet: Keypoint-Guided Fashion Detection".
To avoid problems, please install this repo in a pure conda virtual environment.
First, enter the root directory of this repo. Install CUDA and PyTorch with conda.
conda install -c pytorch -c conda-forge pytorch==1.4.0 torchvision==0.5.0 cudatoolkit-dev=10.1 Then, install other dependencies with pip.
pip install -r requirements.txtcd deepfashion2_api/PythonAPI
pip install -e .Our code is based on mmdetection, which is a clean open-sourced project for benchmarking object detection methods.
cd ../../mmdetection
python setup.py developNow the repo is ready, let's go back to the root directory.
cd ..If you need to run experiments on the entire DeepFashion2 dataset, please refer to DeepFashion2 for detailed guidance. Otherwise, you can skip to the Demo dataset subsection.
After downloading and unpacking the dataset, please create a soft link from the code repository to the dataset's root directory.
ln -s <root dir of DeepFashion2> data/deepfashion2We provide a subset (32 images) of DeepFashion2 to enable quick-experiment.
The checkpoints can be fetched from this OneDrive link.
Test with 1 gpu
./mmdetection/tools/dist_test.sh configs/kgdet_moment_r50_fpn_1x-demo.py checkpoints/KGDet_epoch-12.pth 1 --json_out work_dirs/demo_KGDet.json --eval bbox keypoints- Results files will be stored as
work_dirs/demo_KGDet.json. - If you only need the prediction results, you can drop
--evaland its arguments.
Train with 4 gpus
./mmdetection/tools/dist_train.sh configs/kgdet_moment_r50_fpn_1x-deepfashion2.py 4 --validate --work_dir work_dirs/TRAIN_KGDet- The running log and checkpoints will be stored in the
work_dirs/TRAIN_KGDetdirectory according to the argument--work_dir. --validateevokes a validation section after each training epoch.
Test with 4 gpus
./mmdetection/tools/dist_test.sh configs/kgdet_moment_r50_fpn_1x-deepfashion2.py checkpoints/KGDet_epoch-12.pth 4 --json_out work_dirs/result_KGDet.json --eval bbox keypoints- Results files will be stored as
work_dirs/result_KGDet.json.
If you would like to run our model on your own data, you can imitate the structure of the demo_dataset (an image directory plus a JSON file), and adjust the arguments in the configuration file.
This repo is built upon RepPoints and mmdetection.
@inproceedings{qian2021kgdet,
title={KGDet: Keypoint-Guided Fashion Detection},
author={Qian, Shenhan and Lian, Dongze and Zhao, Binqiang and Liu, Tong and Zhu, Bohui and Li, Hai and Gao, Shenghua},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={35},
number={3},
pages={2449--2457},
year={2021}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent