siyuanliii/masa

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1377 stars · 90 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Official PyTorch implementation of MASA (CVPR 2024 Highlight), a universal instance appearance model that learns to match any objects across video frames using Segment Anything Model outputs, without tracking labels. It includes a universal adapter that adds zero-shot multi-object tracking and segmentation capabilities to existing detection or segmentation models.

Use cases

• track any object in video without domain-specific training labels
• add multi-object tracking to an existing detector like YOLO-World
• segment and track objects jointly in videos with SAM
• zero-shot open-vocabulary multiple object tracking
• evaluate on MOT and MOTS benchmarks like TAO TETA and BDD100K
• learn instance appearance embeddings from unlabeled static images

When to choose

• you need cross-domain or open-vocabulary object tracking without labeled video data
• you want to bolt tracking onto an existing detection or segmentation model
• you are researching instance association and appearance matching

When to avoid

• you need a production-ready turnkey video analytics product
• you lack GPU resources for inference or training
• you need simple single-object tracking only

Facets

library · maturity active

computer-vision machine-learning image-processing deep-learning computer-vision machine-learning artificial-intelligence python object-tracking multiple-object-tracking segment-anything zero-shot-tracking instance-association cvpr-2024 research-code video-analysis linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="siyuanliii/masa")