Questions about Custom Dataset Bounding Box and Voxel Grid Initialization

[yt0409]

thank you for this excellent work! I've successfully run the code and have two questions regarding custom dataset preparation:

When creating my custom DTU-style dataset using COLMAP, I noticed the generated bounding box behaves differently compared to your preprocessed DTU datasets (see attached comparison). Your implementation produces an ideal axis-aligned bounding box, while my COLMAP-based version shows irregular boundaries.
What's the exact methodology for computing the bounding box in your DTU preprocessing pipeline?
Are there recommended parameters or adjustments to achieve similar initialization quality for custom datasets?

The current sphere-in-bbox initialization works well for object-centric scenes. However, for camera-centric 360° capture scenarios (like the one I'm working with), this initialization occurs to bad result. Whether the framework supports alternative initialization strategies for such cases?

[btoussai]

Hello,
For the DTU dataset, I'm using the preprocessed files from Voxurf, itself using preprocessed files from IDR.
The dataset has been normalized to that the object is within a unit sphere, hence I'm using a bounding box from -1 to +1 on all axes. I'm not sure whether this normalization has been computed automatically (from the evaluation pointclouds?) or has been manually adjusted.

The bounding box should be as tight as possible and yet fully encompass the object, with some padding.

I am not very confident that the current algorithm will work for 360° scenes as the setting differs significantly from the typical scenario that we target here (object reconstruction / human reconstruction with background images or masks).

In the last image that you show, the bounding box is very large. This means that some voxels will likely be seen from far away, while others are much closer. This can lead to under-sampling and over-sampling of the grid which is detrimental for convergence, since all voxels have the same size at a given level of detail.

This may not be too much of an issue for indoor scenes, where the distance to the furthest surface is bounded, but will pose problem for unbounded outdoor scenes.

The sphere initialization assumes that the cameras are outside of the sphere, not inside.
You may change this by editing the code here, by flipping the sign of the variable:
https://gitlab.inria.fr/projects-morpheo/ProbeSDF/-/blame/main/resources/shaders/reconstructionVolume/fillVoxelsEllipsoid.cp?ref_type=heads#L47
You can also edit the radius, taken as 0.5 here.
This is a glsl shader so there is no need to recompile the program, just save the file for changes to take effect.

The other type of initialization is the visual hull, that we use for human reconstruction when background images are available, but this is probably not applicable to your case.

A number of approaches specifically deal with 360° scenes and are highly accurate in this setting, you may want to have a look at them.
For instance, see PGSR, GOF (both based on gaussian splatting), or SVRaster.

Regards,
Briac

[yt0409]

thank you for your help! i will try it soon^_^