CrossDepth: Geometry-Constrained Attention for Generalizable Multi-View Surround Depth Estimation

On September 4, 2026, arXiv cs.CV published the CrossDepth method, aimed at addressing the issue of inconsistent depth estimation across images caused by differences in camera parameters and limited field of view of individual images. This method introduces pixel-level camera-perception ray embedding through conditional features to handle parameter differences, and utilizes a geometric-rational region expansion attention mechanism based on calibration devices. The model is trained using photometric consistency in a fully self-supervised manner and verified on the DDAD and nuScenes datasets. Experiments show that compared to state-of-the-art self-supervised methods, CrossDepth improves overall depth accuracy and cross-image depth consistency in both indoor and outdoor evaluations. The related code is available publicly.