CV

SAM 2

SAM 2

Also known as: Segment Anything Model 2, SAM2

SAM 2 (Meta, 2024) is the second-generation Segment Anything Model: prompt with a click, box, or mask and it returns pixel-perfect segmentation, now with native video support that propagates masks across frames automatically. The transformer backbone runs ~6× faster than SAM 1 with better small-object accuracy. FI Tech embeds SAM…

Definition

SAM 2 (Meta, 2024) is the second-generation Segment Anything Model: prompt with a click, box, or mask and it returns pixel-perfect segmentation, now with native video support that propagates masks across frames automatically. The transformer backbone runs ~6× faster than SAM 1 with better small-object accuracy. FI Tech embeds SAM 2 in our annotation tooling (CVAT + custom plugin) so an annotator clicks once on a steel beam and SAM 2 traces the entire object plus tracks it through every frame of a 30-minute drone flyover. Output masks then train downstream YOLO-Seg or RF-DETR-Seg models.