Reproduce “EgoForce: Robust Online Egocentric Motion Reconstruction via Diffusion Forcing” on OpenResearch

About this paper

Inwoo Hwang, Donggeun Lim, Hojun Jang, Young Min Kim

EgoForce introduces an online framework for reconstructing full-body human motion from noisy egocentric observations, adapting diffusion models to operate under strict causal and real-time constraints. It achieves state-of-the-art performance in reconstruction accuracy and motion quality with a latency of 0.08s, demonstrating robustness to sparse and noisy inputs for real-time applications.

Key insights

• Diffusion models, typically offline, can be successfully adapted for strictly causal and online tasks by assigning independent, varying per-frame noise levels and using progressive, limited-step denoising.
• Incorporating causal observation injection and robustness to synthetic noise during training is effective for handling sparse, noisy, and partial egocentric inputs in real-time.
• Maintaining long-term motion coherence in online diffusion inference can be achieved through a stabilization trick that re-injects minimal noise into historical latents to prevent error accumulation.

Summary by alphaXiv.