What is sam-mt?

SAM-MT is a research project from Fudan University and Shanghai University of Finance and Economics that lets a computer track and separate multiple objects in a video at the same time, based on simple clicks from a person. The paper behind it was accepted at a computer vision conference called ECCV in 2026, and the code here is the official release that goes with that paper. The main idea is video segmentation, which means drawing an outline around each object of interest in every frame of a video so a computer knows exactly where that object is over time. Many existing tools slow down a lot as more objects are added to track at once. SAM-MT is built to stay fast even as the number of tracked targets grows, reportedly running at over 36 frames per second while tracking 10 objects at once on a high end NVIDIA graphics card. The repository includes a pretrained model checkpoint that can be downloaded from Hugging Face, along with two ways to try it: a basic script that takes coordinates as input, and an interactive demo built with Gradio, a tool for building simple web interfaces, where a user can click directly on objects in a video to select them. It also includes evaluation scripts for measuring both speed and accuracy against standard video segmentation benchmarks used in this research area. Setting it up requires a computer with Python, the Conda environment manager, and a capable NVIDIA GPU, since the model is designed for real time performance rather than running well on a CPU. Training code is not yet included according to the project's own todo list, only inference and evaluation. This project is aimed at computer vision researchers and engineers working on video object tracking and segmentation, rather than general users, and assumes familiarity with Python, machine learning tooling, and command line workflows.