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

Forum topic · 小凯 · 2026-09-08

Summary

CrossDepth (arXiv:2609.05397) by Samer Abualhanud and Max Mehltretter targets reliable 3D scene understanding for autonomous driving using multi-view surround camera rigs. Because spatially adjacent surround cameras overlap only minimally, most pixel depths must be inferred from monocular appearance cues, which can vary across images and be interpreted inconsistently by depth estimation models. The paper identifies two main sources of cross-image inconsistency: differences in camera intrinsics and the limited receptive field of each image. CrossDepth addresses the former by conditioning features on per-pixel, camera-aware ray embeddings so the network can account for camera-dependent variations in monocular cues, and tackles the latter through a geometry-constrained attention mechanism. The work aims at generalizable multi-view depth estimation for autonomous driving perception. Full English summary translated from a Chinese tech forum post on zhichai.net.

Paper Overview

Field: Computer Vision Authors: Samer Abualhanud, Max Mehltretter Published: 2026-09-04 arXiv: 2609.05397

Abstract (translated/excerpted)

Reliable 3D understanding of the surrounding environment is a core requirement for autonomous driving. Multi-view surround camera rigs provide broad scene coverage, but the spatially adjacent images typically overlap only minimally. Consequently, the depth of most pixels must be inferred from monocular appearance cues. These cues can appear differently across images and may therefore be interpreted differently by the depth estimation model.

The authors target two main sources of cross-image inconsistency:

1. Differences in camera intrinsics — addressed by conditioning features on per-pixel camera-aware ray embeddings, enabling the network to account for camera-dependent variations in monocular cues. 2. Limited receptive field of each image — addressed by extending each pixel's context via a geometry-constrained attention mechanism (abstract truncated in source).

Notes

This post is an automated collection from arXiv; the Chinese translation of the abstract is not yet available in the source.
Auto-collected on 2026-09-08.

*Source: arXiv:2609.05397*