Zero-Shot Evaluation of SAM2 for Surgical Video Segmentation
A systematic benchmark of the promptable video-segmentation foundation model SAM2 applied zero-shot to surgical video across many datasets, surgery types, and imaging modalities. It characterizes how prompting (points, boxes, masks) and finetuning (dense, sparse) strategies affect accuracy, robustness to surgical challenges (tissue deformation, instrument variability), and cross-procedure generalization, finding strong performance in structured scenes but weaknesses in temporal coherence and domain artifacts, and derives guidelines for data-efficient surgical segmentation.
Systematic Evaluation and Guidelines for Segment Anything Model in Surgical Video Analysis
This paper presents the first comprehensive evaluation of the zero-shot capability of SAM2 - a promptable video segmentation foundation model pretrained on natural videos - for surgical video segment…