Scene partitioning for large scale MVS

[Yzhbuaa]

@AIBluefisher, hello, is it possible to partition the scene like CMVS or method introduced by Maura, Massimo et al. using GraphSfM? Should I just use lower completeness ratio when performing point_cloud_segmenter?

I found there is huge overlapping area(e.g. 50% when I set completeness_ratio=0.5) between clusters of my sfm result. Is it possible to use lower completeness_ratio when performing distributed_mapper? I think, perhaps, using lower completeness_ratio will greatly improve the efficiency distributed_mapper, cause there will be less views in each cluster.

Thank you!

[AIBluefisher]

Hi, I don't have much knowledge of CMVS. The point_cloud_segmenter is just served as an experimental tool right now, and I mainly use it for distributed bundle adjustment. I'm now busy on something else, so I feel sorry that I currently don't have much time to revise CMVS. Would you like to show me the main idea of CMVS?

[Yzhbuaa]

Thank you for your response! Sure, as far as I know, the main idea of CMVS and method introduced by Maura, Massimo et al can be illustrated using the following picture: Maura's methods uses camera view information available from Structure-from-Motion (SfM) for computing a set of overlapping clusters suited for Multi-View Stereo (MVS) reconstruction. The main difference between the clusters of GraphSfM and Maura's clusters is that they have different percentage of overlapping area. Specifically, GraphSfM's clusters requires at least 50% of overlapping, while Maura's clusters requires less overlapping. I guess that the key point of huge overlapping area requirement in GraphSfM is to ensure reliable component merging. The key point of Maura's method is to reconstruct dense point cloud with high quality.

• Could GraphSfM's cluster generation method cover the needs of large scale MVS?
• How could I reduce the overlapping area to get better clusters to satisfy the needs of large scale MVS?
• Is it possible to use less overlapping area to achieve the same robustness and higher efficiency in clusters merging step in GraphSfM？

The picture listed above was provided by: Mauro, Massimo, et al. "Overlapping camera clustering through dominant sets for scalable 3D reconstruction." 2013 British Machine Vision Conference