SAMURAI
Zero-shot visual object tracker that adapts SAM 2 with motion-aware memory — no retraining required.
About
SAMURAI is a research project that takes Meta's Segment Anything Model 2 and turns it into a strong visual object tracker without any retraining. SAM 2 already has a memory bank for video, but it was built for general segmentation; SAMURAI adds motion modeling and a smarter memory-selection module that keeps the best past frames to reference. The result tracks objects through occlusion and fast motion far better than vanilla SAM 2, and the authors released the whole thing open source in late 2024.
Independent evaluations back the claim: in the DAM4SAM (CVPR 2025) benchmark table, SAMURAI ranked second overall in quality and robustness, ahead of SAM 2.1 itself, and a construction-site study found it held 92.67% detection accuracy under severe occlusion where YOLOv8 dropped to 52.67%. The trade-off is speed — it runs around 9–12 FPS versus YOLOv8's 28–32, so it's best where accuracy under occlusion matters more than real-time throughput.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Independent eval (DAM4SAM, CVPR 2025): Quality 0.680 / Accuracy 0.722 / Robustness 0.930 — 2nd overall, ahead of SAM 2.1 (0.649/0.720/0.887)
- MDPI construction study: 92.67% vs YOLOv8 52.67% detection accuracy under severe occlusion
- Author VOT evals on LaSOT, GOT-10k, TrackingNet (author-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER metrics, metrics-v1). Evidence: arXiv paper (2411.11922) with VOT benchmark evals (LaSOT, GOT-10k, TrackingNet); official Apache-2.0 repo (yangchris11/samurai) with 7,000+ stars; corroborated by independent DAM4SAM and MDPI evaluations. Strengths: genuinely open, third-party results support the claims. Docked for: (1) headline tracking scores come from the authors' own evaluation protocol; (2) slower than real-time detectors, limiting production use. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · ArXiv 2411.11922; VOT evals on LaSOT/GOT-10k/TrackingNet; official repo yangchris11/samurai, Apache-2.0, 7,000+ stars
- · Independent DAM4SAM (CVPR 2025): 2nd overall, ahead of SAM 2.1; MDPI construction study: 92.67% vs YOLOv8 52.67% under severe occlusion
- · Headline tracking scores author-reported; 9–12 FPS — slower than real-time detectors