BeileiCui/SurgicalDINO

Zero-shot evaluation

yifliu3 opened this issue · 2 comments

yifliu3 commented

Hi, thanks for your great work.

Can you release the code for the inference on SCARED dataset? I'm wondering how you evaluate the relative depth generated by DINOv2.

BeileiCui commented

Hi, thanks for your questions.

I did not prepare well for the code of other parts of current supervised model. But I am going to release another version of Self-Supervised Learning method of depth estimation with more comprehensive codes including training, testing, visualization, 3d reconstruction, trained weights in the near future. So if you are still interested in it, I can recall you to it when I release it.

In terms of evaluating depth on SCARED, we re-scale the predicted depth map by a median scaling method introduced by SfM-Leaner, which can be expressed by:

D_scaled = (D_pred * ( median(D_gt) / median(D_pred))),

where median() refers to find the median, D_gt is the ground truth depth, D_pred is predicted depth, D_scaled is re-scaled depth. Then we evaluate the performance metrics between D_scaled and D_gt.

yifliu3 commented

Thanks for your prompt reply. My problems are solved and looking forward to your codes.

  • 👍1