Researchers achieve 3rd place in video object segmentation challenge using SAM 3 with intelligent re-prompting and DINOv3-based object matching

Developed automatic re-prompting framework built on SAM 3 to handle challenging scenarios like target disappearance, severe transformations, and similar-looking distractors

Method uses DINOv3-based object-level matching with transformation-aware feature pooling to identify and retrieve reliable target anchors from video frames

Achieved 51.17% J&F score on MOSEv2 track test set, ranking 3rd in the PVUW 2026 Challenge for semi-supervised video object segmentation

Enables multi-anchor propagation instead of relying on single initial prompt, improving robustness across multiple core segmentation challenges