Object classification accuracy is too law

[eslambakr]

Dears,

Thank u so much for sharing your amazing work with the codes and checkpoints, this is really appreciated and help us to build above your amazing work.

I have a weird observation during the training the classification loss; obj3d_cls and obj3d_cls_pre, both are too low (almost around 28% as shown in the below image.

Despite, there is a first stage where we train only the classification pipeline (P++ with the classification head) and in this training the obj3d_cls is almost 60%, and we have freezed the P++ so I think it should be around 60% also.

So do u have any justification why it drops from 60% to ~30%?

Thanks in advance!