Depth bleeding issue for COLMAP dense stereo

[zl...@cornell.edu]

Hi,

When I used COLMAP SfM + dense stereo to obtain the dense depth map of images, especially for Internet photos, if I run both photo-metric + geometric runs, in the final depth map, the depth of foreground objects might be propagated to background buildings (i.e. depth bleeding), for example, usually the depth of some regions of foreground trees, statues and street lights would be propagated to the depth of background building. 

It seems that If I only run photo-metric run without geometric filtering, the foreground depths are always not too bad although there are a lot of noise. However, if I run geometric run with geometric filtering, even though I only run 1 or 2 iterations of geometric filtering Patchmatch, there is still serious depth bleeding problem for several images. I am really curious why photo-metric iteration always give better depth for foreground objects without serious depth propagation phenomena than the depth from geometric filtering where serious depth propagation occurs.

Do you think this is because of inaccurate camera poses from SfM or it is the problem of COLMAP dense algorithm itself? Since my goal is to get clean + accurate depth map instead of final 3D models, I really don't want depth of foreground statues propagate to background building even though I can only get semi-dense map. Is there any way workaround to stop depth propagation when I want to run both photo-metric + geometric runs? 

Thanks

[Johannes Schönberger]

Hi,

This is a known limitation of basically all patch-based stereo algorithms. I am currently working on an algorithm to avoid this problem. In the meantime, as a workaround you could reduce the ``window_radius`` option. Hope this explanation helps. Let me know if you have other questions.

Best,

Johannes