Closed pierotofy closed 8 months ago
To reserve the metric scale and camera orientation, it might be easier to feed the original sparse point cloud to training code directly. Rotating the trained ply model file would introduce an extra complicity and quality degradation if it's not handled correctly.
Yes, PlayCanvas's implementation works almost faultlessly. It would be even better if OpenSplat can use the compressed ply format as one of saving options along with the rotation and scaling.
During training the model is scaled and oriented.
Before outputting the output PLY, we should scale / rotate the scene back to its original CRS.