IOU on voxels for SN-Clean not exactly same as your claim in paper about aml-improved-shape-completion HOT 3 CLOSED

From the code, evaluation seems to be the same as in aml/2_b_test.ppy, so I would expect the same results. Are you using the correct model, i.e. the one trained on SN-clean?

Or did you already find the problem?

from aml-improved-shape-completion.

Yes, my stupid mistake. Now for SN-clean, IOU of voxels is exactly same as in your paper.
But for completeness and accuracy measurement, my test result is smaller than in your paper. Did you use the measurement between mesh from predicted sdf and mesh from target sdf in the same resolution, both achieved using marching cube?

from aml-improved-shape-completion.

Sorry for the late reply! No I think, I computed the metrics between the predicted sdf (marching cubes) and the original mesh (not the derived sdf).

from aml-improved-shape-completion.