Event
EgoXperience releases full-stack solution for first-person video reconstruction
1 reports · 1 sources
The EgoXperience team has released a solution covering data collection, automatic annotation, unified modeling, and benchmark evaluation, for jointly reconstructing observer motion, scene geometry, and hand-object interaction from first-person video. Its automatic annotation engine, EgoXX-Engine, combines geometrically verified object tracking with SLAM-based world alignment to produce the 50-hour dataset EgoX-Data-6M (including second-level task descriptions) and the 1-hour evaluation set EgoX-Bench. The team states that, given a monocular first-person video and first-frame object masks, its model can complete joint metric reconstruction of the camera, scene, hands, and objects in a unified world coordinate system in a single forward pass.
Report timeline
- EgoXperience: Learning World-Aligned Egocentric Representations from Human Experience
EgoXperience proposes a complete pipeline covering data collection, automatic annotation, unified modeling, and benchmark evaluation, aimed at jointly reconstructing observer motion, scene geometry, and hand-object interactions from egocentric video. At its core is the automatic annotation engine EgoXX-Engine, which combines geometrically verified object tracking with SLAM-based world alignment to produce the 50-hour dataset EgoX-Data-6M (with second-level task descriptions) and the 1-hour evaluation set EgoX-Bench.
HuggingFace Daily Papers ·