RDVSv2: A Large-scale Benchmark for RGB-D Video Salient Object Detection

Researchers introduced RDVSv2, a new large-scale RGB-D video salient object detection benchmark built from public stereoscopic online videos.

The dataset includes 249 sequences and 29,077 annotated frames with depth maps and eye-tracking-guided salient object masks.

They also adapted SAM2 with parameter-efficient fine-tuning to jointly use RGB, depth, and optical flow.

The resulting baseline achieved state-of-the-art performance on RDVSv2 and other RGB-D VSOD benchmarks.

RDVSv2 offers a higher-quality testbed for evaluating and improving multi-modal video understanding models.