Multi-View Depth Map Estimation With Cross-View Consistency
In Proceedings British Machine Vision Conference 2014
AbstractTo increase robustness of depth-map-based multi-view stereo approaches, we combine several techniques: Depth estimates are propagated in parallel in the local neighborhood to efficiently spread reliable depth information into regions without prominent structures. A faster coarse-to-fine strategy fills in larger holes. Most importantly, a novel cross-view filtering stage based on free-space constraints and variance filtering, enforces consistency among the depth maps of different views. Our algorithm alternates between correlation and consistency optimization. This way, noisy patches and spikes are excluded so that the task for the subsequent depth map fusion algorithms becomes easier. Combining improved propagation, hierarchical estimation, and iterative multi-view consistency optimization, our method increases the estimation speed, generates dense depth maps with desirable global consistency, and yields convincing 3D reconstruction results.
FilesExtended Abstract (PDF, 1 page, 2.9M)
Paper (PDF, 13 pages, 6.2M)
Supplemental Materials (ZIP, 5.5M)