You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Since all DTU scenes share nearly identical intrinsics and the robotic camera arm targeted the same camera poses in each scene yielding the nearly identical extrinsics, does IGEV-MVS training on DTU give scale sensitivity and intrinsic bias?
I'm using 'colmap_input.py' to convert to IterMVS-Net form for IGEV-MVS, which works when starting with DTU input images but yields poor depth maps and very low final mask percentages for custom images, e..g:
As above, is it possible that extrinsics scale could be responsible? In general is there some reference for how the IGEV-MDS DTU model generalizes to custom photo input? Models trained on datasets with limited diversity for camera parameters (including focal length, and principal point) can perform poorly when tested on data with different parameters, a known challenge.
One note: For IGEV-MVS, ‘pair.txt’ must have exactly (10) source images for each reference image. However, the IterMVS-to-Colmap script 'colmap_input.py' exports a source image for every reference image. Therefore I cull the results to exactly (10) source images as per the format specification.
The text was updated successfully, but these errors were encountered:
Since all DTU scenes share nearly identical intrinsics and the robotic camera arm targeted the same camera poses in each scene yielding the nearly identical extrinsics, does IGEV-MVS training on DTU give scale sensitivity and intrinsic bias?
I'm using 'colmap_input.py' to convert to IterMVS-Net form for IGEV-MVS, which works when starting with DTU input images but yields poor depth maps and very low final mask percentages for custom images, e..g:
processing X://dtu//scan400, ref-view00, geo_mask:0.001213 final_mask: 0.001213
As above, is it possible that extrinsics scale could be responsible? In general is there some reference for how the IGEV-MDS DTU model generalizes to custom photo input? Models trained on datasets with limited diversity for camera parameters (including focal length, and principal point) can perform poorly when tested on data with different parameters, a known challenge.
One note: For IGEV-MVS, ‘pair.txt’ must have exactly (10) source images for each reference image. However, the IterMVS-to-Colmap script 'colmap_input.py' exports a source image for every reference image. Therefore I cull the results to exactly (10) source images as per the format specification.
The text was updated successfully, but these errors were encountered: