Recon object not in the object space with SDF-VAE

[ZJU-SJ]

Hi, I am currently endeavoring to train the joint SDF-VAE to acquire SDF modulations. Successful reconstruction is observed when processing data in the object normalization space (as exemplified by ShapeNet). However, when I rotate and translate the object for training, but also normalize the object in the unit cube space. Despite a seemingly well-performing loss, the model struggles to faithfully reconstruct the object. Presented below are the reconstructed results; could this discrepancy be indicative of a limitation in the model?

[GG-Bonds]

Is this the result of how many epochs you trained? My training effect in the first stage was not very good.