ReViV reconstructs egocentric 4D viewer-and-scene dynamics from one monocular video

ethz-vlg · hf · 2026-07-21

ReViV reconstructs both the viewer and the scene in 4D from one egocentric video

ETH Zurich researchers present ReViV, a unified framework for holistic egocentric 4D reconstruction from a single monocular RGB video.

It jointly models RGB video, camera trajectory, gaze, full-body motion, hand motion, and depth as a full multimodal distribution.
The system uses a Masked Generative Egocentric Transformer in a single feed-forward architecture, aiming to avoid separate pipelines for scene perception and ego-motion.
Across HoloAssist, HOT3D, ARCTIC, Aria Digital Twin, and TACO, the paper reports state-of-the-art accuracy and efficiency for ego-body, hand, gaze reconstruction and camera tracking, while staying competitive on depth estimation.
The code and models are open sourced at reviv4d.github.io.

More from Embodied