Segment Anything with Motion, Geometry, and Semantic Adaptation for Complex Nonlinear Visual Object Tracking

Researchers introduced SAMOSA, a SAM 2-based visual object tracker that adds nonlinear motion prediction, semantic recovery, and geometric consistency checks.

The framework is designed to handle occlusion and complex motion more robustly than prior SAM 2-based and supervised tracking methods.

Experiments show especially strong gains on challenging nonlinear motion benchmarks, including anti-UAV datasets.

The work highlights how foundation models can be adapted into stronger, more general-purpose trackers without narrow task-specific supervision.