ByteDance-Seed / ByteDance-Seed/Depth-Anything-3
What does auxiliary feat mean?
- Dominant language
- Python
- Stars
- 6.3k
- Forks
- 702
- PR merge metrics
- No merged PRs in 30d
Description
Good job !!
I can see the potential of upgrade dimension (from 2D to 3D) through your work, which will make a great progress on 3d reconstruction.
Being excited, I read the Depth-Anything-3 code and have some questions that I am curious about:
1, In most of the cases, you have used DINO v2 as the backbone to generate the feature from images. I wonder that have you frozen the backbone while training DA3 model ?
2, When compared to VGGT which is also a DINO based model, DA3 outperforms it and what kind of tricks you have applied when you design the DPT head and perform the training ?
3, Concretely, in da3.py, I still don't quite understand the usage of auxiliary feature since it is not mentioned in your paper and with few comments in your code. Is the auxiliary feature from the distinct layers (determined by the input parameter "export_feat_layers" parameter ) of DINO backbone? If I want to get the features from distinct layers in DPT head, is there an api for me to call ?
Contributor guide
No contributing guide indexed for this repository
Assessment
This issue has not been assessed yet.