search

naver / dust3r

DUSt3R: Geometric 3D Vision Made Easy
https://dust3r.europe.naverlabs.com/
Other
4.65k stars 515 forks source link

Monocular Depth Estimation #82

Closed Davidyao99 closed 2 months ago

Davidyao99 commented 2 months ago

I am trying to use Dust3r to obtain depth maps for a single image. I fed in the same image twice into the model for inference. I was expecting the pts3d from pred1 and pts3d_in_other_view from pred2 to match, but it appears that there is a non-zero difference between the two predictions (difference is of the order 1e-3, with actual values in the order of 1e-1).

Is this expected behavior? And if so, should I just take the output from pred1 as the depth? or average the two results to get a better depth estimation?

Thanks and great work!

yocabon commented 2 months ago

Hi, It's not surprising -> we didn't train it for this. The prediction for img1 is usually more accurate than the one for img2, so it should be better to just take the output from pred1.

Davidyao99 commented 2 months ago

Hello

I see, thank you so much for the quick response!

David