search

lpiccinelli-eth / UniDepth

Universal Monocular Metric Depth Estimation
Other
624 stars 51 forks source link

About smaller V2 models #52

Open wangqiuwen1006 opened 5 months ago

wangqiuwen1006 commented 5 months ago

Dear author, I want to express my appreciation for this project. Your work has been incredibly valuable to me.

I learned that you have plans for releasing smaller v2 versions of the model from README, and I was wondering if you could provide any specific information regarding the timeline for the release of the v2 smaller models. Meanwhile, is there any plan to release the corresponding training code in future updates? I'm looking forward to your reply!

Thank you once again for your efforts and for sharing your work!

lpiccinelli-eth commented 5 months ago

Thank you for your appreciation!

You can check the branch update_v2, there you can load vit-small, the architecture has slightly changed in the last upsampling part (that is why the different branch that will be merged).

wangqiuwen1006 commented 4 months ago

Thank you for your sharing and it is very valuable! I have another small question I would like to ask: I noticed that you used affine-invariant depth data in the training of the v2 model. How did you transform the GT metric depth in the corresponding datasets into scale-shift invariant depth to optimize the network (in other words, how did you define the mathematical formulas for scale and shift)?