MoSA: Segmentation via Video Kinematics

rohanpaul_ai · x · 2026-07-15

This post summarizes MoSA (Motion-Grounded Segment Anything), an ECCV 2026 paper from the 360 AI Research Institute. It falls under the "See Precisely" initiative, aiming to help AI understand visual content more accurately.

The core concept: instead of manually annotating objects one by one, the model learns object boundaries and segmentation information from motion cues in videos. The authors state that MoSA used about 10,000 hours of unlabeled video to generate over 21 million pseudo-labels.

Described as more than just a segmentation method, this work serves a broader research goal: scaling AI perception, editing, and generation capabilities in a more precise and controllable manner.

Related event: MoSA: 360 AI Research learns object segmentation from unlabeled video(5 posts)→

Original post →

More from Research

Research channel →