OpenMVS vs texrecon

[satyajitghana]

I tried openMVS texturing and also texrecon https://github.com/nmoehrle/mvs-texturing

It seems texrecon is much better in terms of quality of texture generated. openMVS creates a large patch of texture by default and tries to pack all the views in that. in texrecon each view is projected individually and separate texture images for each view is saved. which does generate a large number of file is views is more, but the final model looks much better.

It would be nice if texrecon is merged with openMVS texturing pipeline.

[cdcseacave]

not sure what you mean exactly, can you pls share the same model (preferably starting from the same SfM solution) reconstructed in both OpenMVS and texrecon?

[ghost]

Ammonite | 102 Images 6000x4000

TextureMesh -i scene_dense_mesh.mvs --resolution-level 0 --outlier-threshold 0.078 --patch-packing-heuristic 0
  --decimate 0.05 --virtual-face-images 3
    Time: 6m 30s
    Atlases: 1
  --decimate 0.01 --virtual-face-images 3
    Time: 1m 52s
    Atlases: 1
texrecon
  Decimation 1
    Time: 7m 55s
    Atlases: 55
  Decimation 0.05
    Time: 40s
    Atlases: 5

TextureMesh 0.05 vf3

texrecon 0.05

texrecon 1

[satyajitghana]

Ammonite | 102 Images 6000x4000

TextureMesh -i scene_dense_mesh.mvs --resolution-level 0 --outlier-threshold 0.078 --patch-packing-heuristic 0
  --decimate 0.05 --virtual-face-images 3
    Time: 6m 30s
    Atlases: 1
  --decimate 0.01 --virtual-face-images 3
    Time: 1m 52s
    Atlases: 1
texrecon
  Decimation 1
    Time: 7m 55s
    Atlases: 55
  Decimation 0.05
    Time: 40s
    Atlases: 5

TextureMesh 0.05 vf3 texrecon 0.05 texrecon 1

look at the smudges in TextureMesh

These are much better in texrecon

[cdcseacave]

this is a great comparison, thank, however why do you use the virtual faces feature? it is work in progress; or at least compare also to the defualt (mature) algorithm, to be fair, right?

[ghost]

True. I used virtual faces to give TextureMesh its best shot at quality.

TextureMesh
  --decimate 0.05
    Time: 2m 51s
    Atlases: 1

TextureMesh
  --decimate 0.05 --cost-smoothness-ratio 1
    Time: 3m 12s
    Atlases: 1

TextureMesh
  --decimate 0.05 --cost-smoothness-ratio 1 --outlier-threshold 0.0599999987
    Time: 3m 14s
    Atlases: 1

TextureMesh
  --decimate 0.05 --virtual-face-images 3 --outlier-threshold 0.0599999987
    Time: 5m 28s
    Atlases: 1

TextureMesh
  --decimate 0.05 --virtual-face-images 3 --cost-smoothness-ratio 1 --outlier-threshold 0.0599999987
    Time: 5m 11s
    Atlases: 1

[cdcseacave]

very interesting, it seems that each param has quite some influence on the final result when it should not, like the --outlier-threshold for ex
next it would be interesting in identifying the weak link, if it is only one,
anybody willing to help with this?