DART: Depth-as-Target Pretraining for Surgical Vision Foundation Models
2026-09-07 12:00Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
2mentions
SummaryAI generated
On September 7, 2026, arXiv cs.CV released the DART method, which aims to enhance surgical visual base models by introducing depth as a pre-training objective. Based on the DINOv2 architecture, this method adds a pixel-space depth reconstruction task during the pre-training phase to utilize pseudo-label depth signals, while retaining RGB single-modal input during the inference phase. Experimental results show that compared to alternative signals such as edge information, depth information can more effectively encode scene geometry, and the pixel-level reconstruction head improves rather than undermines the model’s representation quality. In eight surgical benchmark tests involving segmentation, depth estimation, and image recognition, the DART method performed better than natural image pre-training baselines and existing benchmarks in the field, achieving both dense prediction and image understanding enhancements.
DART proposes a pre-training method based on depth as the target, utilizing pseudo-label depth signals to improve the performance of surgical vision base models. This method adds pixel-space depth reconstruction targets on top of DINOv2, which is used only during the pre-training phase; inference remains in RGB single modality. Experiments show that depth information can more effectively encode scene geometry than alternative signals such as edges, and this pixel-level reconstruction head improves rather than undermines representation quality. In eight surgical benchmark tests involving segmentation, depth estimation, and image recognition, DART outperforms natural image pre-training baselines and same-domain benchmarks, achieving dual enhancements in dense prediction and image understanding.