[arXiv]score: 0.48
BED-SAM2 Adds Monocular Depth to SAM2 Encoder, Improves Camouflaged Object Segmentation
May 26, 2026
BED-SAM2 modifies SAM2's Hiera encoder to incorporate monocular depth from RGB images, achieving competitive state-of-the-art on salient and camouflaged object detection in as few as five training epochs.
cs.CV
HOW THIS AFFECTS YOU
●
builderFive-epoch fine-tuning to competitive SOTA on camouflaged object detection makes BED-SAM2 a practical starting point for segmentation applications requiring geometric awareness.
●
researcherInjecting monocular geometric priors into SAM2's encoder is a lightweight adaptation strategy that improves boundary delineation for camouflaged objects with minimal training.