Use computed detections/keypoints to prompt a segmentation model

[sfmig]

Given:

• a video, and
• a set of detections (bboxes) per frame, or keypoints per frame.

Use the detections (or keypoints) to prompt a segmentation model (such as SAM2) to generate tracked segmentation masks.

Consider using SAMURAI, which seems to deal better with error propagation in zero-shot performance