Use Record3d exported dataset for training Gaussian Splatting

[tungts1101]

Is there any tutorial or script that helps me extract the exported data from Record3D to replace the estimated camera pose/initial point cloud from COLMAP, so that I can use it later for training Gaussian Splatting model?

[marek-simonik]

There is no tutorial/script per-se, but you can take a look at #93 (comment) to see how the exported .r3d files are structured (the comment contains links to other Github issues where the camera pose format is described in more detail).

You can also take a look at how nerfstudio parses Record3D's metadata[.json] file: https://github.com/nerfstudio-project/nerfstudio/blob/main/nerfstudio/process_data/record3d_utils.py

As an alternative to exporting your 3D videos into .r3d, it might be more practical to export into EXR + JPG instead — you will get an EXR (containing float32 depth in meters) and a JPG (containing RGB images) file per frame, and there will also be the metadata.json file with camera poses (the same metadata as those in the .r3d files).