Integral Human Pose Regression

The major contributors of this repository include Xiao Sun, Chuankang Li, Bin Xiao, Fangyin Wei, Yichen Wei.

Introduction

Integral Regression is initially described in an ECCV 2018 paper. (Slides).

We build a 3D pose estimation system based mainly on the Integral Regression, placing second in the ECCV2018 3D Human Pose Estimation Challenge. Note that, the winner Sarandi et al. also uses the Integral Regression (or soft-argmax) with a better augmented 3D dataset in their method indicating the Integral Regression is the currently state-of-the-art 3D human pose estimation method.

The Integral Regression is also known as soft-argmax. Please refer to two contemporary works (Luvizon et al. and Nibali et al.) for a better comparision and more comprehensive understanding.

Disclaimer

This is an official implementation for Integral Human Pose Regression based on Pytorch. It is worth noticing that:

• The original implementation is based on our internal Mxnet version. There are slight differences in the final accuracy and running time due to the plenty details in platform switch.

License

© Microsoft, 2017. Licensed under an MIT license.

Citing Papers

If you find Integral Regression useful in your research, please consider citing:

@article{sun2017integral,
  title={Integral human pose regression},
  author={Sun, Xiao and Xiao, Bin and Liang, Shuang and Wei, Yichen},
  journal={arXiv preprint arXiv:1711.08229},
  year={2017}
}

@article{sun2018integral,
  title={An Integral Pose Regression System for the ECCV2018 PoseTrack Challenge},
  author={Sun, Xiao and Li, Chuankang and Lin, Stephen},
  journal={arXiv preprint arXiv:1809.06079},
  year={2018}
}

Main Results

Train on CHALL_H80K train, evaluate on CHALL_H80K val.

description	box det.	dataset	backbone	patch size	flip test	ensemble	MPJPE(mm)	model
original baseline	No	HM36	ResNet-50	-	No	No	115.9
+person box det.	Yes	HM36	ResNet-50	256*256	No	No	86.5(25.4%)
+MPII data	Yes	HM36+MPII	ResNet-50	256*256	No	No	62.2(28.1%)
+deeper	Yes	HM36+MPII	ResNet-152	256*256	No	No	61.2(1.6%)
+larger image	Yes	HM36+MPII	ResNet-152	288*384	No	No	58.5(4.4%)
+COCO data	Yes	HM36+MPII+COCO	ResNet-152	288*384	No	No	57.5(1.7%)	download
+flip test	Yes	HM36+MPII+COCO	ResNet-152	288*384	Yes	No	56.9(1.0%)	download
+model ensemble	Yes	HM36+MPII+COCO	ResNet-152	288*384	Yes	Yes	55.3(2.8%)	download

Train on CHALL_H80K train+val, evaluate on CHALL_H80K test.

description	box det.	dataset	backbone	patch size	flip test	ensemble	MPJPE(mm)	model
challenge result	Yes	HM36+MPII+COCO	ResNet-152	288*384	Yes	Yes	47	download

Environment

Python Version: 3.6
OS: CentOs7 (Other Linux system is also OK)
CUDA: 9.0 (least 8.0)
PyTorch:0.4.0(see issue #4)

Installation

We recommend installing python from Anaconda, installing pytorch following guide on PyTorch according to your specific CUDA & python version. In addition, you need to install dependencies below.

pip install scipy
pip install matplotlib
pip install opencv-python
pip install easydict
pip install pyyaml

Preparation for Training & Testing

1. Download Human3.6M(ECCV18 Challenge) image from Human3.6M Dataset and our processed annotation from Baidu Disk (code: kfsm) or Google Drive
2. Download MPII image from MPII Human Pose Dataset
3. Download COCO2017 image from COCO Dataset
4. Download cache file from Dropbox
5. Organize data like this

${PROJECT_ROOT}
 `-- data
     `-- coco
        |-- images
        |-- annotations
        |-- COCO_train2017_cache
     `-- mpii
        |-- images
        |-- annot
        |-- mpii_train_cache
        |-- mpii_valid_cache
     `-- hm36
        |-- images
        |-- annot
        |-- HM36_train_cache
        |-- HM36_validmin_cache
     `-- hm36_eccv_challenge
        `-- Train
            |-- IMG
            |-- POSE
        `-- Val
            |-- IMG
            |-- POSE
        `-- Test
            |-- IMG
        |-- HM36_eccv_challenge_Train_cache
        |-- HM36_eccv_challenge_Test_cache
        |-- HM36_eccv_challenge_Val_cache

Usage

We have placed some example config files in experiments folder, and you can use them straight forward. Don't modify them unless you know exactly what it means.

Train

For Integral Human Pose Regression, cd to pytorch_projects/integral_human_pose
Integral Regression

python train.py --cfg=experiments/hm36/resnet50v1_ft/d-mh_ps-256_deconv256x3_min-int-l1_adam_bs32-4gpus_x300-270-290/lr1e-3.yaml --dataroot=../../data/  

Direct Joint Regression

python train.py --cfg=experiments/hm36/resnet50v1_ft/d-mh_ps-256_dj_l1_adam_bs32-4gpus_x140-90-120/lr1e-3.yaml --dataroot=../../data/

For 3D pose estimation system of ECCV18 Challenge, cd to pytorch_projects/hm36_challenge

python train.py --cfg=experiments/hm36/resnet152v1_ft/d-mh_ps-256_deconv256x3_min-int-l1_adam_bs24-4gpus_x300-270-290/lr1e-3.yaml --dataroot=../../data/

By default, logging and model will be saved to log and output folder respectively.

Test

To run evaluation on CHALL_H80K Val dataset

1. Download model
2. Place it under $project_root/model/hm36_challenge
3. cd to $project_root/pytorch_projects/hm36_challenge
4. execute command below

python test.py --cfg experiments/hm36/resnet152v1_ft/d-mch_384x288_deconv256x3_min-int-l1_adam_bs12-4gpus/lr1e-4_x300-270-290.yaml --model=../../model/hm36_challenge/model_chall_train_152ft_384x288.pth.tar

Extensions

• The project is built on old version of pytorch(0.4.0), and currently the latest released one has updated to 1.0.1. So there may be some compatibility problems. Please feel free to submit new issues.
• A third-party implementation by mks0601
• A third-party implementation by mkocabas