Our paper has been accepted as Spotlight by ECCV2020
This is a pytorch realization of Residual Steps Network which won 2019 COCO Keypoint Challenge and ranks 1st place on both COCO test-dev and test-challenge datasets as shown in COCO leaderboard. The original repo is based on the inner deep learning framework (MegBrain) in Megvii Inc.
In this paper, we propose a novel method called Residual Steps Network (RSN). RSN aggregates features with the same spatialsize (Intra-level features) efficiently to obtain delicate local representations, which retain rich low-level spatial information and result in precise keypoint localization. In addition, we propose an efficient attention mechanism - Pose Refine Machine (PRM) to further refine the keypoint locations. Our approach won the 1st place of COCO Keypoint Challenge 2019 and achieves state-of-the-art results on both COCO and MPII benchmarks, without using extra training data and pretrained model. Our single model achieves 78.6 on COCO test-dev, 93.0 on MPII test dataset. Ensembled models achieve 79.2 on COCO test-dev, 77.1 on COCO test-challenge dataset. The source code is publicly available for further research.
Model | Input Size | GFLOPs | AP | AP50 | AP75 | APM | APL | AR |
---|---|---|---|---|---|---|---|---|
Res-18 | 256x192 | 2.3 | 70.7 | 89.5 | 77.5 | 66.8 | 75.9 | 75.8 |
RSN-18 | 256x192 | 2.5 | 73.6 | 90.5 | 80.9 | 67.8 | 79.1 | 78.8 |
RSN-50 | 256x192 | 6.4 | 74.7 | 91.4 | 81.5 | 71.0 | 80.2 | 80.0 |
RSN-101 | 256x192 | 11.5 | 75.8 | 92.4 | 83.0 | 72.1 | 81.2 | 81.1 |
2×RSN-50 | 256x192 | 13.9 | 77.2 | 92.3 | 84.0 | 73.8 | 82.5 | 82.2 |
3×RSN-50 | 256x192 | 20.7 | 78.2 | 92.3 | 85.1 | 74.7 | 83.7 | 83.1 |
4×RSN-50 | 256x192 | 29.3 | 79.0 | 92.5 | 85.7 | 75.2 | 84.5 | 83.7 |
4×RSN-50 | 384x288 | 65.9 | 79.6 | 92.5 | 85.8 | 75.5 | 85.2 | 84.2 |
Model | Input Size | GFLOPs | AP | AP50 | AP75 | APM | APL | AR |
---|---|---|---|---|---|---|---|---|
RSN-18 | 256x192 | 2.5 | 71.6 | 92.6 | 80.3 | 68.8 | 75.8 | 77.7 |
RSN-50 | 256x192 | 6.4 | 72.5 | 93.0 | 81.3 | 69.9 | 76.5 | 78.8 |
2×RSN-50 | 256x192 | 13.9 | 75.5 | 93.6 | 84.0 | 73.0 | 79.6 | 81.3 |
4×RSN-50 | 256x192 | 29.3 | 78.0 | 94.2 | 86.5 | 75.3 | 82.2 | 83.4 |
4×RSN-50 | 384x288 | 65.9 | 78.6 | 94.3 | 86.6 | 75.5 | 83.3 | 83.8 |
4×RSN-50+ | - | - | 79.2 | 94.4 | 87.1 | 76.1 | 83.8 | 84.1 |
Model | Input Size | GFLOPs | AP | AP50 | AP75 | APM | APL | AR |
---|---|---|---|---|---|---|---|---|
4×RSN-50+ | - | - | 77.1 | 93.3 | 83.6 | 72.2 | 83.6 | 82.6 |
Model | Split | Input Size | Head | Shoulder | Elbow | Wrist | Hip | Knee | Ankle | Mean |
---|---|---|---|---|---|---|---|---|---|---|
4×RSN-50 | val | 256x256 | 96.7 | 96.7 | 92.3 | 88.2 | 90.3 | 89.0 | 85.3 | 91.6 |
4×RSN-50 | test | 256x256 | 98.5 | 97.3 | 93.9 | 89.9 | 92.0 | 90.6 | 86.8 | 93.0 |
Model | Input Size | GFLOPs | AP | AP50 | AP75 | APM | APL | AR |
---|---|---|---|---|---|---|---|---|
Res-18 | 256x192 | 2.3 | 65.2 | 87.3 | 71.5 | 61.2 | 72.2 | 71.3 |
RSN-18 | 256x192 | 2.5 | 70.4 | 88.8 | 77.7 | 67.2 | 76.7 | 76.5 |
- + means using ensemble models.
- All models are trained on 8 V100 GPUs
- We done all the experiments using our own DL-Platform MegDL, all results in our paper are reported on MegDL. There are some differences between MegDL and Pytorch. MegDL will be released in March. The MegDL code and model will be also publicly avaible.
This repo is organized as following:
$RSN_HOME
|-- cvpack
|
|-- dataset
| |-- COCO
| | |-- det_json
| | |-- gt_json
| | |-- images
| | |-- train2014
| | |-- val2014
| |
| |-- MPII
| |-- det_json
| |-- gt_json
| |-- images
|
|-- lib
| |-- models
| |-- utils
|
|-- exps
| |-- exp1
| |-- exp2
| |-- ...
|
|-- model_logs
|
|-- README.md
|-- requirements.txt
-
Install Pytorch referring to Pytorch website.
-
Clone this repo, and config RSN_HOME in /etc/profile or ~/.bashrc, e.g.
export RSN_HOME='/path/of/your/cloned/repo'
export PYTHONPATH=$PYTHONPATH:$RSN_HOME
- Install requirements:
pip3 install -r requirements.txt
- Install COCOAPI referring to cocoapi website, or:
git clone https://github.com/cocodataset/cocoapi.git $RSN_HOME/lib/COCOAPI
cd $RSN_HOME/lib/COCOAPI/PythonAPI
make install
-
Download images from COCO website, and put train2014/val2014 splits into $RSN_HOME/dataset/COCO/images/ respectively.
-
Download ground truth from Google Drive or Baidu Drive (code: fc51), and put it into $RSN_HOME/dataset/COCO/gt_json/.
-
Download detection result from Google Drive or Baidu Drive (code: fc51), and put it into $RSN_HOME/dataset/COCO/det_json/.
-
Download images from MPII website, and put images into $RSN_HOME/dataset/MPII/images/.
-
Download ground truth from Google Drive or Baidu Drive (code: fc51), and put it into $RSN_HOME/dataset/MPII/gt_json/.
-
Download detection result from Google Drive or Baidu Drive (code: fc51), and put it into $RSN_HOME/dataset/MPII/det_json/.
Create a directory to save logs and models:
mkdir $RSN_HOME/model_logs
Go to specified experiment repository, e.g.
cd $RSN_HOME/exps/RSN50.coco
and run:
python config.py -log
python -m torch.distributed.launch --nproc_per_node=gpu_num train.py
the gpu_num is the number of gpus.
python -m torch.distributed.launch --nproc_per_node=gpu_num test.py -i iter_num
the gpu_num is the number of gpus, and iter_num is the iteration number you want to test.
Please considering citing our projects in your publications if they help your research.
@misc{cai2020learning,
title={Learning Delicate Local Representations for Multi-Person Pose Estimation},
author={Yuanhao Cai and Zhicheng Wang and Zhengxiong Luo and Binyi Yin and Angang Du and Haoqian Wang and Xinyu Zhou and Erjin Zhou and Xiangyu Zhang and Jian Sun},
year={2020},
eprint={2003.04030},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@article{li2019rethinking,
title={Rethinking on Multi-Stage Networks for Human Pose Estimation},
author={Li, Wenbo and Wang, Zhicheng and Yin, Binyi and Peng, Qixiang and Du, Yuming and Xiao, Tianzi and Yu, Gang and Lu, Hongtao and Wei, Yichen and Sun, Jian},
journal={arXiv preprint arXiv:1901.00148},
year={2019}
}
@inproceedings{chen2018cascaded,
title={Cascaded pyramid network for multi-person pose estimation},
author={Chen, Yilun and Wang, Zhicheng and Peng, Yuxiang and Zhang, Zhiqiang and Yu, Gang and Sun, Jian},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={7103--7112},
year={2018}
}
And the code of Cascaded Pyramid Network is also available.
You can contact us by email published in our paper or [email protected].