Question about Training Details

[sspsk]

Dear Authors,
Congrats on your CVPR acceptance!

In the forward method of the model I see there is the option of not passing joint/motion parameters. So, during training do you also use non articulated objects (to train only for part segmentation)?

In preliminary training experiments with particulate with ~3.800 articulated objects and using the same augmentation techniques, the model suffers from overfitting for the part segmentation problem as indicated from my 'point_mask_loss' validation curves. Would be helpful if you have any suggestions to this matter.

Thanks in advance,
Sotiris

[RuiningLi]

Hi Sotiris, thanks for your interest.

Yes we opened this as an option but during training for our released checkpoint we didn't use any generic objects with part annotations only -- but I'm bullish in large-scale pretraining as a first step to boost the model's generalization to novel categories!