Aggregation and Finetuning for Clothes Landmark Detection

[WIP] Code release is still in preparation, stay tuned.

1st place solution for CVPR 2020 DeepFashion2 Clothes Landmark Detection Track. This repo contains code for the keypoints detector. The object detection part is not contained.

Prerequisite

Install

install HRNet dependencies as in HRNet/HRNet-Human-Pose-Estimation, then
```
cd HRNet-Human-Pose-Estimation/lib
make
```

install cocoapi in switchablenorms/DeepFashion2

git clone https://github.com/switchablenorms/DeepFashion2
cd DeepFashion2/deepfashion2_api/PythonAPI/
make install

Data preprocess

download data to HRNet-Human-Pose-Estimation/lib/data/deepfashion2/

preprocess data to following structure

data/deepfashion2/
	annotations/ # original data, converted from deepfashion2_to_coco.py
		train_coco.json
		validation_coco.json
		test_coco.json
	annotations_agg81kps/ # for training in aggregation
		train_coco_agg81kps.json
	annotations_per_category/ # for validation in finetuning
		validation_coco_category_1.json
		...
		validation_coco_category_13.json
	annotations_per_category_agg81kps/ # for training in finetuning
		train_coco_category_1.json
		...
		train_coco_category_13.json
	images/ # images, make some symlinks (or just cp) for this to be done
		train/
		validation/
		test/

COCO pretrained weights is downloaded from here to HRNet-Human-Pose-Estimation/models/pytorch/pose_coco/pose_hrnet_w48_384x288.pth

Usage

Training

Aggregation

python tools/train.py --cfg experiments/deepfashion2/w48_512x384_adam_lr1e-3-agg81kps.yaml

Finetune

# finetuning 1st category
python tools/train.py --cfg experiments/deepfashion2/w48_512x384_adam_lr1e-3-agg81kps-category1-hflip.yaml
# finetuning 2nd category
python tools/train.py --cfg experiments/deepfashion2/w48_512x384_adam_lr1e-3-agg81kps-category2-hflip.yaml
# finetuning others ...

Testing

Testing on validation set (with gt det)

CUDA_VISIBLE_DEVICES=0,1,2,3 python tools/test.py \
    --cfg experiments/deepfashion2/w48_512x384_adam_lr1e-3-agg81kps-category1-hflip.yaml \
GPUS '(0,1,2,3)' \
    TEST.MODEL_FILE output/deepfashion2agg81kps/pose_hrnet/w48_512x384_adam_lr1e-3-agg81kps-category1-hflip/final_state.pth \
    TEST.USE_GT_BBOX True \
    TEST.IMAGE_THRE 0.01 \
    DATASET.TEST_SET validation \
    TEST.FLIP_TEST True

Testing on validation set (with object det results /path/to/det.json)

CUDA_VISIBLE_DEVICES=0,1,2,3 python tools/test.py \
    --cfg experiments/deepfashion2/w48_512x384_adam_lr1e-3-agg81kps-category1-hflip.yaml \
GPUS '(0,1,2,3)' \
    TEST.MODEL_FILE output/deepfashion2agg81kps/pose_hrnet/w48_512x384_adam_lr1e-3-agg81kps-category1-hflip/final_state.pth \
    TEST.USE_GT_BBOX False \
    TEST.COCO_BBOX_FILE '/path/to/det.json' \
    TEST.IMAGE_THRE 0.01 \
    DATASET.TEST_SET validation \
    TEST.FLIP_TEST True

...

Reference

HRNet/HRNet-Human-Pose-Estimation

Please cite following if you use the code

@misc{lin2020aggregation,
    title={Aggregation and Finetuning for Clothes Landmark Detection},
    author={Tzu-Heng Lin},
    year={2020},
    eprint={2005.00419},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

License

MIT

lzhbrian/deepfashion2-kps-agg-finetune