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

Last update: Dec 23, 2022

Overview

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

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

Install dependencies.

pip install -r requirements.txt

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}
}

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

Related tags

Overview

TransFusion-Pose

Overview

Installation

Data preparation

Human 3.6M

Ski-Pose

Training and Testing

Human 3.6M

Ski-Pose

Citation

Acknowledgement

Owner

Haoyu Ma

Implementation of TimeSformer, a pure attention-based solution for video classification

PyTorch implementation DRO: Deep Recurrent Optimizer for Structure-from-Motion

Official implementation of "Learning Proposals for Practical Energy-Based Regression", 2021.

My personal code and solution to the Synacor Challenge from 2012 OSCON.

PyTorch implementation of Lip to Speech Synthesis with Visual Context Attentional GAN (NeurIPS2021)

A Streamlit demo demonstrating the Deep Dream technique. Adapted from the TensorFlow Deep Dream tutorial.

SmallInitEmb - LayerNorm(SmallInit(Embedding)) in a Transformer to improve convergence

PyTorch code for the "Deep Neural Networks with Box Convolutions" paper

A simplistic and efficient pure-python neural network library from Phys Whiz with CPU and GPU support.

Reproducing Results from A Hybrid Approach to Targeting Social Assistance

Predict bus arrival time using VertexAI and Nvidia's Jetson Nano

PyTorch implementation of "Learning to Discover Cross-Domain Relations with Generative Adversarial Networks"

Introducing neural networks to predict stock prices

CoCosNet v2: Full-Resolution Correspondence Learning for Image Translation

Build upon neural radiance fields to create a scene-specific implicit 3D semantic representation, Semantic-NeRF

Nest Protect integration for Home Assistant. This will allow you to integrate your smoke, heat, co and occupancy status real-time in HA.

Variational autoencoder for anime face reconstruction

This is the code of using DQN to play Sekiro .

Complete the code of prefix-tuning in low data setting

DSTC10 Track 2 - Knowledge-grounded Task-oriented Dialogue Modeling on Spoken Conversations