Grounded-SAM-2

Why we included this project

Grounded SAM 2 pairs open-vocabulary detectors like Grounding DINO and Florence-2 with SAM 2's segmentation and tracking engine. You describe an object in plain language, and the pipeline masks it and follows it across frames, which makes it a practical choice for automated video annotation, object counting, and scene analysis when you want something promptable instead of detectors built around a fixed set of classes. The authors deliberately simplified the code relative to the original research implementation, so it also reads well as an example of combining detection, segmentation, and tracking in one flow, with the supervision library handling the visualization. Tiled inference is available for high-resolution imagery, which helps teams working with large frames.