Table of Contents

Paper: SAM 3: Segment Anything with Concepts (Meta, ICLR 2026).

The one-line idea

Earlier Segment Anything models answer “segment the thing I clicked.” SAM 3 answers “segment every instance of this concept” — given a short noun phrase like “pressure gauge” or an example image, it returns masks for all matches and tracks them through video.

Numbers worth remembering

• On the SA-Co/Gold open-vocabulary benchmark: 54.1 cgF1, versus 24.6 for OWLv2.
• LVIS instance segmentation: 48.8 mask AP (DINO-X: 38.5).
• About 30 ms per image with 100+ objects on an H200.

Limits the authors are honest about

• Prompts are simple noun phrases; “the second gauge from the left” needs an external multimodal LLM.
• Specialized domains still benefit from fine-tuning.

Why I care (robotics view)

Robot maps store points, not objects. A concept-level segmenter plus synchronized depth lets us label points by name — remove every person from a map, or keep every gauge as an inspection target. The open question is how well this holds on small, industrial, long-tail objects seen from a moving robot.