SAM-MT
Real-Time Interactive Multi-Target Video Segmentation

1Fudan University
2Shanghai University of Finance and Economics

ECCV 2026

Interactive Demo

SAM-MT Interactive Demo

SAM-MT supports click-based real-time multi-target video segmentation regardless of target count.

Abstract

Modern video object segmentation tracks and segments user-specified targets. While recent approaches such as SAM2 achieve strong single-target performance, extending them to multi-target scenarios usually requires repeating object-wise propagation for each target, causing latency and memory usage to grow with target count. We present SAM-MT, a real-time interactive multi-target video segmentation framework built upon SAM2. SAM-MT represents the scene with shared global context and represents each target with lightweight explicit queries. It uses decoupled masked attention to prevent cross-target interference, a query-based sparse memory to preserve target identities over time, and an identity transformer for robust target re-identification. This design largely decouples latency from the number of targets, enabling real-time multi-target segmentation with robust video segmentation performance.

Architecture

SAM-MT converts SAM2 from object-wise propagation into a query-driven multi-target framework. Dense memory (of SAM2) is used only once to model the combined global context, while each individual target is represented by a compact target query. In the decoder, decoupled masked attention allows all target queries to access the shared global queries, but blocks direct attention between different targets. Across frames, historical target queries are stored in a FIFO sparse memory, and an identity transformer updates each target only from its own history to avoid identity drift.

SAM-MT framework overview
Target Queries

Explicit target identity

Each user-specified target is represented by a lightweight query, avoiding repeated object-wise SAM2 propagation.

Decoupled Masked Attention

Identity-preserving attention

Different targets are isolated from each other to reduce cross-target interference, while still sharing global context.

Sparse Memory

Efficient temporal modeling

Queries are stored into sparse memory instead of pixel-level features, reducing per-target overhead for scalability.

Quantitative Results

SAM-MT achieves competitive video segmentation performance across VOS benchmarks. More importantly, it maintains real-time speed as the number of targets increases, such as sustaining 35+ FPS with 20 targets with low VRAM overhead.

VOS Performance

SAM-MT quantitative results on VOS benchmarks

FPS vs. Number of Targets

SAM-MT FPS scalability comparison

VRAM vs. Number of Targets

SAM-MT VRAM scalability comparison

Qualitative Results

SAM-MT tracks and segments many targets simultaneously while preserving individual identities in crowded scenes.

SAM-MT qualitative results in dense multi-target scenes

BibTeX

@inproceedings{SAMMT,
  title={{SAM-MT}: Real-Time Interactive Multi-Target Video Segmentation},
  author={Shen, Ruiqi and Liu, Chang and Ding, Henghui},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}