integral-human-pose
The major contributors of this repository include Xiao Sun, Chuankang Li, Bin Xiao, Fangyin Wei, Yichen Wei.
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.
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.
© Microsoft, 2017. Licensed under an MIT license.
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} }
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 |
Python Version: 3.6
OS: CentOs7 (Other Linux system is also OK)
CUDA: 9.0 (least 8.0)
PyTorch:0.4.0(see issue https://github.com/JimmySuen/integral-human-pose/issues/4)
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
Download Human3.6M(ECCV18 Challenge) image from Human3.6M Dataset and our processed annotation from Baidu Disk (code: kfsm) or Google Drive
Download MPII image from MPII Human Pose Dataset
Download COCO2017 image from COCO Dataset
Download cache file from Dropbox
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
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.
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.
To run evaluation on CHALL_H80K Val dataset
Download model
Place it under $project_root/model/hm36_challenge
cd to $project_root/pytorch_projects/hm36_challenge
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
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.
上一篇:Face-Pose-Net
还没有评论,说两句吧!
热门资源
Keras-ResNeXt
Keras ResNeXt Implementation of ResNeXt models...
seetafaceJNI
项目介绍 基于中科院seetaface2进行封装的JAVA...
spark-corenlp
This package wraps Stanford CoreNLP annotators ...
capsnet-with-caps...
CapsNet with capsule-wise convolution Project ...
inferno-boilerplate
This is a very basic boilerplate example for pe...
智能在线
400-630-6780
聆听.建议反馈
E-mail: support@tusaishared.com