MoSA: 360 AI Research learns object segmentation from unlabeled video

A paper from 360 AI Research, MoSA (Motion-Grounded Segment Anything), drew focused discussion across several posts at ECCV 2026. The work argues that models should learn object concepts and segmentation boundaries by observing motion in unlabeled video, freeing them from dependence on large-scale manually annotated images. It attracted attention because it directly confronts the SAM family's heavy reliance on annotated data.

Key details

Posts broadly describe MoSA as a method that learns "visual concepts from motion": rather than having humans outline objects frame by frame, the model uses motion cues in video to judge which regions form independent objects. According to the summaries, the method uses roughly 10,000 hours of unlabeled video as training material, and falls under a "See Precisely" direction aimed at letting AI see more precisely and form visual concepts by observing the world.

Background and significance

Several authors benchmark against Meta's SAM: while SAM can segment arbitrary objects, it relies on hundreds of millions of manually annotated masks. MoSA's value is therefore framed as building the ability to "segment anything" on more scalable video observation rather than costly manual labor. The current posts mostly stay at the level of the paper's idea and positioning, without offering concrete experimental metrics or effect details.

2026-07-14 ~ 2026-07-15 · 5 related posts

• [source] MoSA Learns Object Segmentation from Unlabeled Video — heyshrutimishra · 2026-07-14
• MoSA: Segmentation via Video Kinematics — rohanpaul_ai · 2026-07-15
• 360 Research: Object Segmentation via Video Kinematics — eyishazyer · 2026-07-15

2 near-duplicate retellings: rohanpaul_ai · ZabihullahAtal