search

lpiccinelli-eth / UniDepth

Universal Monocular Metric Depth Estimation
Other
462 stars 39 forks source link

Is there a large loss in the x direction? #47

Closed red-liu closed 1 month ago

red-liu commented 1 month ago

I found there is a larger error in the x direction. Is the model only used to predict depth, or is there a special configuration for the model?

red-liu commented 1 month ago

predicted intrinsic value: fx=1471.4409, fy=2411.942, cx=1517.1675, cy=2021.4147 The values of predicted intrinsic fx and fy are quite different, is there a problem?

lpiccinelli-eth commented 1 month ago

I doubt there is a problem with the inference. I remember some new datasets (like MegaDepth or similar) that are COLMAP-based and present "weird" focal lengths for some scenes, especially f_x. It could be that your input image is similar to those MegaDepth scenes and the model just memorized those "weird" covariates, ie (image, depth, camera), since they are pretty off and there is not much data to hedge those "weird" covariates.

Anyway, thank you for pointing this out, we will dive deeper into this problem. One simple (temporary) solution on your side is to use only f_y.

red-liu commented 1 month ago

I appreciate your detailed answer regarding my question. thanks your great work.