OmniDS: Dual-Stream Context Fusion for Omnidirectional Depth from Fisheye Cameras

2026-07-03

AI Summary

This work addresses the challenges of omnidirectional depth estimation in multi-fisheye camera systems, where wide baselines induce inter-view observation inconsistencies, leading to ambiguous feature fusion near occlusion boundaries and fine structures. To resolve this, the authors propose OmniDS, an iterative depth refinement framework that decouples context construction from cross-view fusion. OmniDS employs a dual-stream encoder (CNN combined with a frozen DINOv3 backbone), a consensus-aware multi-view cost volume, and a geometric distortion-biased deformable cross-attention mechanism to enable dynamic, geometry-aware feature reprojection in equirectangular projection (ERP) space. Integrated with 3D U-Net regularization and knowledge distillation into a lightweight MobileNet architecture, OmniDS achieves state-of-the-art accuracy on the OmniThings, OmniHouse, and Sunny benchmarks while maintaining efficient inference.

Abstract

Omnidirectional depth estimation from multi-fisheye camera rigs is complicated by visibility conflicts: wide baselines cause different cameras to observe different portions, or even different faces, of the same object, so aggregating their features into a unified equirectangular (ERP) representation under fixed projection produces ambiguous matching evidence near occlusion boundaries and thin structures. Although existing methods mitigate this by down-weighting unreliable views, they do not resolve the underlying discrepancy because context formation and cross-view fusion remain tied to rigid fisheye-to-ERP sampling. We present OmniDS, an iterative depth refinement framework that replaces rigid aggregation by combining dynamic context fusion with consensus-aware multi-view similarity. A dual-stream encoder pairs a lightweight CNN for geometric detail with a frozen DINOv3 for semantic priors; their features are reprojected into ERP space at each refinement step via learned view weighting and deformable cross-attention with geometric distortion bias. In parallel, a multi-view consensus volume captures global cross-camera agreement through group-wise correlation and feature variance, regularized by a 3D U-Net. For efficient deployment, we distill the dual-stream representation into a single MobileNet-based encoder. OmniDS achieves state-of-the-art performance on the OmniThings, OmniHouse, and Sunny benchmarks while maintaining competitive inference speed. Project page and codes are available at https://parkchaesong.github.io/omnids.

Problem

• omnidirectional depth estimation
• fisheye cameras
• visibility conflicts
• equirectangular projection
• multi-view fusion

Innovation

• dual-stream fusion
• dynamic context aggregation
• multi-view consensus
• deformable cross-attention
• omnidirectional depth estimation