Metric-Guided Feature Fusion of Visual Foundation Models for Segmentation Tasks

In three linesMetric-guided fusion approach combining complementary features from visual foundation models (SAM2, DINOv3) for dense prediction tasks. Two label-free metrics (Structural Coherence, Edge Fidelity) assess encoders and select complementary pairs. Consistent performance gains across multiple tasks without complex architectural changes.

Summary generated by Claude — human-verified