Difference in Results When Running ProbesDF on DTU Dataset After COLMAP Reconstruction

[wanghzen0808]

Hello,
I am encountering an issue when using the DTU dataset with ProbesDF. When I directly use the dataset, ProbesDF produces great results. However, after running COLMAP on the dataset images and then running ProbesDF, the results are not as good as the original DTU dataset.

Could you please help me understand why this might be happening? Are there specific settings or steps I need to follow in COLMAP to maintain the quality of the results when running ProbesDF?

Thank you for your assistance!

[btoussai]

Hello,
first of all I'm glad to hear you managed to download and run the code.

If I understand correctly, you are trying to use colmap to calibrate the cameras instead of using the calibration that is already included with the dataset. There are several possibilities that can explain the gap in quality.

1. It could be that colmap wasn't able to compute a calibration which is as accurate as the one provided in the dataset, in that case there is nothing I can do.
2. Another possibility is that you haven't been able to convert colmap's data format to the one we are using. In that case the results should be very bad instead of slightly bad.
3. The calibration has been correctly converted but the position and orientation of the reconstruction volume (the scene's bounding box) is wrong. In that case the results could be slightly bad if the voxel size (relative to the cameras) is a bit different to what is used by default, or very bad is the voxel size is very different or if the reconstruction volume does not even include the object.

You can check calibration issue using the graphical interface. You will be able to visualize the position of the cameras with respect to the reconstruction volume (the blue wireframe box), as well as to look at the images (for instance some images may be swapped due to colmap listing the cameras in a different order, this will completely mess up the result as well).

Remember that colmap will produce a calibration in a completely different reference frame than the one used originally. This means that you must adjust the rotation, translation and scale of the reconstruction volume and the size of the voxels. If you cancel this transformation correctly, there shouldn't be any observable difference in the user interface between what colmap produces and the official calibration (assuming colmap managed to properly calibrate the cameras).

The information that you gave is very limited so it is difficult to be more specific than that.
I hope my explanations are clear.
Best,
Briac