3AM: 3egment Anything with Adaptive Geometric Consistency in Videos
3AM introduces a training-time enhancement to SAM2 that integrates 3D-aware features from MUSt3R to achieve geometry-consistent recognition. This allows for robust object tracking across drastic viewpoint changes using only RGB input, eliminating the need for explicit camera poses or expensive 3D ground truth masks. The method fuses visual appearance with implicit geometric correspondence, enabling stable tracking even when objects move significantly or disappear temporarily.