Semantic-aware Transfer with Instance-adaptive Parsing for Crowded Scenes Pose Estimation

This is an official implementation of Semantic-aware Transfer with Instance-adaptive Parsing for Crowded Scenes Pose Estimation

This repo is built on deep-high-resolution-net.

Main Results

Results on CrowdPose test set

Method	Backbone	Input size	AP	Ap@50	AP@75	AP (Easy)	AP (Medium)	AP (Hard)
HRNet	HRNet-w32	256 x 192	71.7	89.8	76.9	79.6	72.7	61.5
HRNet + STIP	HRNet-w32	256 x 192	74.1	90.0	79.9	81.6	75.1	64.3
HRNet	HRNet-w48	256 x 192	73.3	90.0	78.7	81.0	74.4	63.4
HRNet + STIP	HRNet-w48	256 x 192	74.8	90.8	80.1	82.0	75.7	65.0

Results on COCO val set

Method	Backbone	Input size	AP	Ap@50	AP@75	AP (Small)	AP (Medium)	AP (Large)
HRNet	HRNet-w32	256 x 192	74.4	90.5	81.9	70.8	81.0	79.8
HRNet + STIP	HRNet-w32	256 x 192	75.8	90.3	82.4	72.1	82.4	80.8
HRNet	HRNet-w48	256 x 192	75.1	90.6	82.2	71.5	81.8	80.4
HRNet + STIP	HRNet-w48	256 x 192	76.0	90.4	82.2	72.2	82.9	81.1

Installation

The environment can be referred to README.md.

The details about dataset can be referred to README.md.

Downlaod pretrained weights from BaidunYun(Password: cr1x) to ./models.

Testing on CrowdPose dataset using pretrained weights

Testing HRNet

python tools/script_test.py \
    --cfg experiments/crowdpose/hrnet/w32_256x192_adam_lr1e-3.yaml \
    TEST.MODEL_FILE models/pytorch/crowdpose/hrnet_w32_256x192.pth

Testing STIP net

python tools/script_test.py \
    --cfg experiments/crowdpose/partnet/stipnet_w32_256x192_adam_lr1e-3.yaml \
    TEST.MODEL_FILE models/pytorch/crowdpose/stipnet_w32_256x192.pth

Training on CrowdPose dataset

python tools/script_train.py \
    --cfg experiments/crowdpose/partnet/stipnet_w32_256x192_adam_lr1e-3.yaml 

Citation

If you find this work or code is helpful in your research, please cite:

@inproceedings{pose:stip,
	author={Xuanhan Wang and Lianli Gao and Yan Dai and Yixuan Zhou and Jingkuan Song},
	title={Semantic-aware Transfer with Instance-adaptive Parsing for Crowded Scenes Pose Estimation},
	pages={686--694},
	booktitle = {ACM MM},
	year={2021}
}