Official topic - 3D reconstruction and rendering reconstructed 3D meshes

[COLIBRI VR]

Hi,

This is the official topic for discussing 3D reconstruction and rendering reconstructed 3D meshes.

Feel free to post!

--

Grégoire Dupont de Dinechin

[Guillaume Brincin]

Hi,

I have been working with colibri VR for more than a week.

First thank you for making your work available.

I have been testing the datasets with the recommended pipeline colmap -> blender -> instant mesh -> rendering as explained in your last tutorial video. Most of the time the results are satisfaying.

However once I try to use my own datasets the 3d reconstruction is really poor (part missing, really rough mesh) resulting in very poor rendering. As those datasets give great results in reconstruction software such as metashape.

Do you have some "good practice"  advice for image capturing? Is there a way to configure colmap to have a more precise reconstructed mesh?

Thank you for your answer.

Best Regards

[Greg de Dinechin]

Hi,

Thank you very much for your interest in the project! And very sorry about the late reply, I hadn't been keeping up to date with the forum.

There is definitely an issue with the current way we implemented the link to external tools, in that it lacks the flexibility that enables tuning the reconstruction parameters to the specificities of the dataset. That being said, if you are familiar with other reconstruction tools such as Metashape (I've also recently started working with Meshroom, which I recommend), the simplest solution may be to export or convert the resulting 3D mesh in the OBJ format, and select this mesh during rendering with COLIBRI VR. The issue here is then to also export the positions, orientations, and intrinsic parameters of the cameras in the COLMAP format (i.e. cameras.txt and images.txt files), but this should be possible as well, depending on how Metashape enables you to export these parameters. You would also need to export the undistorted images and place them in the "images" folder.

The other option is to find a way to get similar results with COLMAP, but I agree that the results are sometimes far less convincing for some datasets. For advice on image capture for 3D reconstruction, I recommend such guides as Unity's Photogrammetry workflow, or dedicated blog posts such as this one. As for configuring COLMAP, it very much depends on the images, and I keep returning to the COLMAP documentation myself to try to understand how to better handle some datasets, so my best advice would be to try to work with the documentation and see if it has any tips that may work for you.

I hope this answers your questions, feel free to keep me updated!

Best,

Greg