HowieMa / TransFusion-Pose

[BMVC2021] "TransFusion: Cross-view Fusion with Transformer for 3D Human Pose Estimation"

TransFusion-Pose

TransFusion: Cross-view Fusion with Transformer for 3D Human Pose Estimation
Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin and Xiaohui Xie
In BMVC 2021
[Paper] [Video]

Overview

• We propose the TransFusion, which apply the transformer architecture to multi-view 3D human pose estimation
• We propose the Epipolar Field, a novel and more general form of epipolar line. It readily integrates with the transformer through our proposed geometry positional encoding to encode the 3D relationships among different views.
• Extensive experiments are conducted to demonstrate that our TransFusion outperforms previous fusion methods on both Human 3.6M and SkiPose datasets, but requires substantially fewer parameters.

Installation

1. Clone this repo, and we'll call the directory that you cloned multiview-pose as ${POSE_ROOT}

git clone https://github.com/HowieMa/TransFusion-Pose.git

2. Install dependencies.

pip install -r requirements.txt

3. Download TransPose models pretrained on COCO.

wget https://github.com/yangsenius/TransPose/releases/download/Hub/tp_r_256x192_enc3_d256_h1024_mh8.pth

You can also download it from the official website of TransPose

Please download them under ${POSE_ROOT}/models, and make them look like this:

${POSE_ROOT}/models
└── pytorch
    └── coco
        └── tp_r_256x192_enc3_d256_h1024_mh8.pth

Data preparation

Human 3.6M

For Human36M data, please follow H36M-Toolbox to prepare images and annotations.

Ski-Pose

For Ski-Pose, please follow the instruction from their website to obtain the dataset.
Once you download the Ski-PosePTZ-CameraDataset-png.zip and ski_centers.csv, unzip them and put into the same folder, named as ${SKI_ROOT}.
Run python data/preprocess_skipose.py ${SKI_ROOT} to format it.

Your folder should look like this:

${POSE_ROOT}
|-- data
|-- |-- h36m
    |-- |-- annot
        |   |-- h36m_train.pkl
        |   |-- h36m_validation.pkl
        |-- images
            |-- s_01_act_02_subact_01_ca_01 
            |-- s_01_act_02_subact_01_ca_02

|-- |-- preprocess_skipose.py
|-- |-- skipose  
    |-- |-- annot
        |   |-- ski_train.pkl
        |   |-- ski_validation.pkl
        |-- images
            |-- seq_103 
            |-- seq_103

Training and Testing

Human 3.6M

# Training
python run/pose2d/train.py --cfg experiments-local/h36m/transpose/256_fusion_enc3_GPE.yaml --gpus 0,1,2,3

# Evaluation (2D)
python run/pose2d/valid.py --cfg experiments-local/h36m/transpose/256_fusion_enc3_GPE.yaml --gpus 0,1,2,3  

# Evaluation (3D)
python run/pose3d/estimate_tri.py --cfg experiments-local/h36m/transpose/256_fusion_enc3_GPE.yaml

Ski-Pose

# Training
python run/pose2d/train.py --cfg experiments-local/skipose/transpose/256_fusion_enc3_GPE.yaml --gpus 0,1,2,3

# Evaluation (2D)
python run/pose2d/valid.py --cfg experiments-local/skipose/transpose/256_fusion_enc3_GPE.yaml --gpus 0,1,2,3

# Evaluation (3D)
python run/pose3d/estimate_tri.py --cfg experiments-local/skipose/transpose/256_fusion_enc3_GPE.yaml

Our trained models can be downloaded from here

Citation

If you find our code helps your research, please cite the paper:

@inproceedings{ma2021transfusion,
  title={TransFusion: Cross-view Fusion with Transformer for 3D Human Pose Estimation},
  author={Ma, Haoyu and Chen, Liangjian and Kong, Deying and Wang, Zhe and Liu, Xingwei and Tang, Hao and Yan, Xiangyi and Xie, Yusheng and Lin, Shih-Yao and Xie, Xiaohui},
  booktitle={British Machine Vision Conference},
  year={2021}
}

Acknowledgement

• Cross-view Fusion
• TransPose
• Epipolar Transformer