Player not loading? Watch on YouTube
This tutorial explains how to use SAM 2 for video object segmentation and tracking in Label Studio. It starts with a labeling configuration using VideoVector and VideoVectorLabels. The presenter defines labels and explains that SAM 2 segments objects from user input rather than label names, so the names can suit the annotation task.
The setup uses the existing Segment Anything 2 video backend linked in the description. The presenter instructs viewers to host the model somewhere with GPU availability, then connect its backend URL through the project's model integrations settings. The connection also needs interactive pre-annotations enabled before validation and saving. This is a hosted model workflow; the tutorial does not provide hardware specifications or walk through backend installation.
For the demonstration, the presenter selects a human label in a soccer clip, enables the box tool and auto annotation, and draws a box around one person. They say points should also work, but prefer a box for this task. After segmentation, the interface provides forward or backward tracking. Because the example starts on the first frame, only forward tracking is available. The timeline fills with segmentation results as tracking progresses.
