[Feature Request] improve signal-to-noise of surface geometry near edges about visfd

[jewettaij]

Today I posted a new version of the code which implements idea number 2 from the post above. I wrote:

"Unfortunately, the oriented point-clouds generated by "filter_mrc" do not look noticeably better after this change. (...when viewed visually in meshlab. See the comment in issue #11, for why.) However the number of points on the surface and the length of their surface normals have been adjusted and normalized in this new version of the code. (This is not obvious if you only look at the point cloud in mesh lab.) As a result, if closed-surface reconstruction (a.k.a. "Screened Poisson Reconstruction", "PoissonRecon") is applied to this new point cloud, there will be a modest but somewhat noticeable reduction in bumpiness in the resulting closed surface near the old surface edges."

After analyzing the consequences of this change, I noticed that most of this benefit probably comes simply from weighting surface normal contributions by saliency (which I was not doing earlier), ...not from shifting the positions of the voxels.

Since I failed to adequately shift the voxel positions in this version of the code, I think this idea can be explored further: Currently, only voxels that are already in the target cluster are considered when searching along each curve. Since these tend to be only 1 voxel wide, it means that most nearby voxels are not explored, even if there is a wide blob of voxels with high saliency located there. A more radical approach would consider all voxels whose saliency is higher than some cutoff, and closer than some distance, regardless of whether they belong to this cluster. This could potentially shift voxel positions by an even greater distance (hopefully for the better).

I can explore this idea later. Now, the new code works at least as well as the code from before version v0.27.3. Since I think I have undone most of the damage that I did in my more recent commits, I will close the issue for now.