Mark12Ding/SAM2Long

[ICCV 2025] SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree

What it solves

SAM2Long addresses the issue of error accumulation in long-term video segmentation. In standard SAM 2, tracking objects across long videos often leads to drift or failure when objects are occluded or reappear, causing the segmentation mask to degrade over time.

How it works

It introduces a training-free memory tree that manages multiple segmentation hypotheses. Instead of relying on a single path of memory, the system dynamically prunes less optimal paths as the video progresses, reducing the risk of error propagation and improving resilience against occlusions.

Who it’s for

Computer vision researchers and developers working on video object segmentation (VOS), particularly those needing high-accuracy tracking in long-form content where objects frequently disappear and reappear.

Highlights

  • Training-Free: Enhances SAM 2 without requiring additional parameters or further training.
  • Resilient Tracking: Specifically designed to handle object occlusion and reappearance in long videos.
  • Performance Boost: Demonstrates significant improvements in J & F scores across benchmarks like SA-V and LVOS compared to the original SAM 2.
  • Compatible: Works with both SAM 2 and SAM 2.1 checkpoints across various backbone sizes (Tiny, Small, Base+, Large).

Related

  • Project
    bytedance/Sa2VA

    Sa2VA is a ByteDance research codebase that fuses SAM‑2 segmentation with multimodal LLMs (InternVL, Qwen‑VL) to provide pixel‑level grounded understanding for images and videos. It includes the core Sa2VA model, the VRT reasoning benchmark, the SAMTok mask‑token interface, and extensions like SaSaSa2VA. The repo uses the `uv` package manager for reproducible environments, provides setup scripts, and links to papers, model zoo, and datasets.

  • Project
    facebookresearch/sam3

    SAM 3 (Segment Anything with Concepts) is Meta’s 848 M‑parameter foundation model that lets you prompt an image or video with free‑form text (or visual exemplars) and receive masks, boxes, and scores for *all* matching objects. It combines a DETR‑style detector and a SAM 2‑style tracker, introduces a presence token for fine‑grained prompt discrimination, and is trained on >4 M auto‑annotated concepts. The repo provides installation steps, example notebooks, and a new SA‑CO benchmark (270 K concepts) for evaluation.

  • Project
    yatengLG/ISAT_with_segment_anything

    An interactive semi-automatic image segmentation annotation tool that uses the Segment Anything Model (SAM) to accelerate the creation of labeled datasets.

  • Project
    Zarxrax/Sammie-Roto-2

    A cross-platform desktop application for AI-assisted video masking, segmentation, and object removal using models like SAM2 and MatAnyone.

  • Project
    kadirnar/segment-anything-video

    A packaged version of the Segment Anything Model (SAM) that provides easy pip installation and extended support for video segmentation and cloud GPU integration.