TRI-ML/DDAD

About the challenge.

jakey-ahn opened this issue · 4 comments

Hi, I'm participating in dense depth challenge, self-supervised track.
Can I use ImageNet pre-trained weight for the model for this competition?
I'm aware of the rule that external data should be only monocular videos, but I want to know if the ImageNet weight is available or not.
Thanks!

Thank you for bringing that up, that's a very good question.
We have decided to allow ImageNet pre-trained weights, since it is standard in most depth estimation works. That is the only exception though, so other sources of pre-training are still not allowed.

@VitorGuizilini-TRI How about using a pretrained semantic segmentation model? This is also commonly used in previously self-supervised methods.

Although interesting (especially given our previous work and inclusion of per-class depth metrics), unfortunately the use of semantic information has been ruled out for this challenge in particular.

@VitorGuizilini-TRI Thanks for the clarification! Can you also answer my question about the extrinsics in another thread?