Pixelwise View Selection for Unstructured Multi-View Stereo

Abstract

This work presents a Multi-View Stereo system for robust and efficient dense modeling from unstructured image collections.

Ourcore contributions are the joint estimation of depth and normal information , pixelwise view selection using photometric and geometric priors , and a multi-view geometric consistency term for the simultaneous refinement and image-based depth and normal fusion.

本文核心贡献是深度和法向信息的联合估计，使用光度和几何先验的像素级视图选择，以及用于同时细化和基于图像的深度和法向融合的多视图几何一致性项。

Experiments on benchmarks and large-scale Internet photo collections demonstrate stateof-the-art performance in terms of accuracy, completeness, and efficiency.

Figure

Figure 1

Reconstructions for Louvre , Todai-ji , Paris Opera , and Astronomical Clock.

Figure 2

(a) Illustration of geometric priors for reference view (R) and three source views (1–3). View 1 has similar resolution (red), and good triangulation (green) and incident angle (blue), while view 2 is oblique and has lower resolution. View 3 cannot see the patch.

(b) Geometric prior likelihood functions with different parameters. (Color figure online)

Figure 3

Reconstruction results for South Building and Fountain.

Figure 4

Photometric and geometric priors for South Building dataset between reference image (R) and each two selected source images (1–5).

Figure 5

(a) Comparison of spatial smoothness term with our proposed spatial and temporal smoothness term for the occlusion variables Z . Algorithm starts from the left with the first sweep and is followed by consecutive sweeps to the right.

(b) Estimated depths and normals using standard PatchMatch propagation (cf.Fig.3 for ours).

(c) Reference image with filtered depths and normals for crowd-sourced images

Figure 6

Reference image with filtered depths and normals for crowd-sourced images.

Conclusion

This work proposes a novel algorithm forrobust and efficientdense reconstruction from unstructured image collections.

Our method estimates accurate depth and normal information using photometric and geometric information for pixelwise view selection and for image-based fusion and filtering.

We achieve stateof-the-art results on benchmarks and crowd-sourced data.