Meta Segment Anything Model 3
7 days ago
- #Meta
- #AI
- #Computer Vision
- Meta introduces Segment Anything Model 3 (SAM 3), featuring text and exemplar prompts for advanced image and video segmentation.
- SAM 3 supports open-vocabulary noun phrases and image exemplars, overcoming limitations of fixed-label segmentation models.
- Meta releases SAM 3 model weights, a demo on Segment Anything Playground, and a research paper detailing its development.
- SAM 3 achieves a 2x performance gain over existing systems on the SA-Co benchmark for promptable concept segmentation.
- A hybrid AI-human data engine accelerates annotation, creating a training set with over 4 million unique concepts.
- SAM 3's architecture leverages Meta Perception Encoder and DETR, achieving fast inference and state-of-the-art performance.
- Applications include Facebook Marketplace's View in Room feature and wildlife monitoring with the SA-FARI dataset.
- Future improvements could enhance SAM 3's efficiency in video processing and complex scene understanding.
- The Segment Anything Playground offers a user-friendly platform for experimenting with SAM 3's capabilities.
- Meta encourages the AI community to adopt SAM 3 and the SA-Co benchmark to advance open science and innovation.