EventEgo3D: 3D Human Motion Capture from Egocentric Event Streams [CVPR'24]

Christen Millerdurai^1,2, Hiroyasu Akada¹, Jian Wang¹, Diogo Luvizon¹, Christian Theobalt¹, Vladislav Golyanik¹

¹ Max Planck Institute for Informatics, SIC         ² Saarland University, SIC

Official PyTorch implementation

Project page | Paper

Abstract

Monocular egocentric 3D human motion capture is a challenging and actively researched problem. Existing methods use synchronously operating visual sensors (e.g. RGB cameras) and often fail under low lighting and fast motions, which can be restricting in many applications involving head-mounted devices. In response to the existing limitations, this paper 1) introduces a new problem, i.e. 3D human motion capture from an egocentric monocular event camera with a fisheye lens, and 2) proposes the first approach to it called EventEgo3D (EE3D). Event streams have high temporal resolution and provide reliable cues for 3D human motion capture under high-speed human motions and rapidly changing illumination. The proposed EE3D framework is specifically tailored for learning with event streams in the LNES representation, enabling high 3D reconstruction accuracy. We also design a prototype of a mobile head-mounted device with an event camera and record a real dataset with event observations and the ground-truth 3D human poses (in addition to the synthetic dataset). Our EE3D demonstrates robustness and superior 3D accuracy compared to existing solutions across various challenging experiments while supporting real-time 3D pose update rates of 140Hz.

Advantages of Event Based Vision

High Speed Motion	Low Light Performance

EventEgo3D

Usage

---

• EventEgo3D: 3D Human Motion Capture from Egocentric Event Streams [CVPR'24]
  • Official PyTorch implementation
    • Abstract
    • Advantages of Event Based Vision
    • EventEgo3D
  • Usage
    • Installation
      • Dependencies
      • Pretrained Model
    • Datasets
    • Training
    • Evaluation
      • EE3D-S
      • EE3D-R
  • Citation
  • License
  • Acknowledgements

---

Installation

Clone the repository

git clone https://github.com/Chris10M/EventEgo3D.git
cd EventEgo3D

Dependencies

Create a conda enviroment from the file

conda env create -f EventEgo3D.yml

Next, install ocam_python using pip

pip3 install git+https://github.com/Chris10M/ocam_python.git

Pretrained Model

The pretrained model best_model_state_dict.pth can be found here. Please place the model in the following folder structure.

EventEgo3D
|
└── saved_models
         |
         └── best_model_state_dict.pth

Datasets

The datasets can obtained by executing the files in dataset_scripts. For detailed information, refer here.

Training

For training, ensure EE3D-S, EE3D-R and EE3D[BG-AUG] are present. The batch size and checkpoint path can be specified with the following environment variables, BATCH_SIZE and CHECKPOINT_PATH.

python train.py 

Evaluation

EE3D-S

For evaluation, ensure EE3D-S Test is present. Please run,

python evaluate_ee3d_s.py 

The provided pretrained checkpoint gives us an accuracy of,

Arch	Head_MPJPE	Neck_MPJPE	Right_shoulder_MPJPE	Right_elbow_MPJPE	Right_wrist_MPJPE	Left_shoulder_MPJPE	Left_elbow_MPJPE	Left_wrist_MPJPE	Right_hip_MPJPE	Right_knee_MPJPE	Right_ankle_MPJPE	Right_foot_MPJPE	Left_hip_MPJPE	Left_knee_MPJPE	Left_ankle_MPJPE	Left_foot_MPJPE	MPJPE	Head_PAMPJPE	Neck_PAMPJPE	Right_shoulder_PAMPJPE	Right_elbow_PAMPJPE	Right_wrist_PAMPJPE	Left_shoulder_PAMPJPE	Left_elbow_PAMPJPE	Left_wrist_PAMPJPE	Right_hip_PAMPJPE	Right_knee_PAMPJPE	Right_ankle_PAMPJPE	Right_foot_PAMPJPE	Left_hip_PAMPJPE	Left_knee_PAMPJPE	Left_ankle_PAMPJPE	Left_foot_PAMPJPE	PAMPJPE
EgoHPE	29.020	44.788	51.028	98.415	144.434	53.148	103.996	141.923	91.309	146.183	210.144	224.728	87.292	141.563	210.710	219.027	124.857	50.010	47.727	50.475	86.859	131.602	53.526	90.658	127.851	74.953	98.317	116.803	129.032	75.543	96.434	119.562	132.058	92.588

EE3D-R

For evaluation, ensure EE3D-R is present. Please run,

python evaluate_ee3d_r.py 

The provided pretrained checkpoint gives us an accuracy of,

Arch	walk_MPJPE	crouch_MPJPE	pushup_MPJPE	boxing_MPJPE	kick_MPJPE	dance_MPJPE	inter. with env_MPJPE	crawl_MPJPE	sports_MPJPE	jump_MPJPE	MPJPE	walk_PAMPJPE	crouch_PAMPJPE	pushup_PAMPJPE	boxing_PAMPJPE	kick_PAMPJPE	dance_PAMPJPE	inter. with env_PAMPJPE	crawl_PAMPJPE	sports_PAMPJPE	jump_PAMPJPE	PAMPJPE
EgoHPE	70.881	163.840	97.886	136.571	103.724	88.877	103.191	109.714	101.020	97.320	107.302	52.113	99.483	75.530	104.667	86.055	71.968	70.859	77.949	77.827	80.179	79.663

Citation

If you find this code useful for your research, please cite our paper:

@inproceedings{Millerdurai_EventEgo3D_2024, 
    title={EventEgo3D: 3D Human Motion Capture from Egocentric Event Streams}, 
    author={Christen Millerdurai and Hiroyasu Akada and Jian Wang and Diogo Luvizon and Christian Theobalt and Vladislav Golyanik}, 
    booktitle = {Computer Vision and Pattern Recognition (CVPR)}, 
    year={2024} 
} 

License

EventEgo3D is under CC-BY-NC 4.0 license. The license also applies to the pre-trained models.

Acknowledgements

The code is partially adapted from here.