Hi, thanks for the great work.
Thanks for the great work, and for releasing the model. I had a question about the output representation of the object voxelization in the Geometry model.
In my tests, the Geometry model does not always seem to output the shape in a canonical pose aligned to fixed axes (e.g., with the longest axis along Z). Moreover, I evaluated the refinement stage model using voxelizations with arbitrary scales and rotations and, as long as the voxelizations are smooth, the output meshes are consistently very good.
I know you are not planning to release the training scripts, but could you please clarify whether this behavior was explicitly encouraged during training (for example, by providing the refinement stage with rotated/scaled smooth voxelizations as input), or whether you trained only on canonical voxelizations and this robustness is an inherent capability of the model?
Thanks again
Hi, thanks for the great work.
Thanks for the great work, and for releasing the model. I had a question about the output representation of the object voxelization in the Geometry model.
In my tests, the Geometry model does not always seem to output the shape in a canonical pose aligned to fixed axes (e.g., with the longest axis along Z). Moreover, I evaluated the refinement stage model using voxelizations with arbitrary scales and rotations and, as long as the voxelizations are smooth, the output meshes are consistently very good.
I know you are not planning to release the training scripts, but could you please clarify whether this behavior was explicitly encouraged during training (for example, by providing the refinement stage with rotated/scaled smooth voxelizations as input), or whether you trained only on canonical voxelizations and this robustness is an inherent capability of the model?
Thanks again