Segment Anything Models

Segment Anything research studies how promptable models can produce object masks from inputs such as points, boxes, text, or prior masks, and how to extend them to video and 3D settings. Common approaches adapt frozen encoders with lightweight adapters or fine-tuning, improve prompts and task-specific decoders, or add temporal memory for tracking and video segmentation. These models are applied to settings including medical imaging, microscopy, remote sensing, and robotics, where limited labels and fine boundaries make reliable segmentation valuable. A central tension is how to preserve broad zero-shot generalization while improving domain-specific accuracy, prompt reliability, temporal stability, and inference efficiency.