Multiple View Stereo with quadtree-guided priors

Multi-View Stereo (MVS) algorithms rely on common photometric consistency measures and, therefore, in cases of low-textured surfaces tend to generate unreliable depth estimates or lack completeness due to matching ambiguities. Such textureless areas often imply dominant planar structures, typically occurring in man-made scenes. To support depth estimation in scenarios where challenging surfaces are present, we propose an extended PatchMatch pipeline using an adaptive accumulated matching cost calculation based on estimated prior plane hypotheses and the local textureness. Plane priors are detected in the object space and guided by quadtree structures in order to generate depth and normal hypothesis for every pixel, supporting, in this way, the propagation of more reliable depth estimates across the image. Experiments on the ETH3D high-resolution dataset and on custom real-world scenes demonstrate that our approach can favor the reconstruction of problematic regions by adding small complexity while preserving fine details in rich textured regions, achieving thus competitive results compared to state-of-the-art methods. The source code of the developed method is available at https://github.com/3DOM-FBK/openMVS.

Automated summary

To support depth estimation in challenging surfaces scenarios, we propose an extended PatchMatch pipeline using an adaptive accumulated matching cost calculation. Our approach achieves competitive results compared to state-of-the-art methods by favoring the reconstruction of problematic regions while preserving fine details in rich textured regions.