search

zhyever / PatchFusion

[CVPR 2024] An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation
https://zhyever.github.io/patchfusion/
MIT License
958 stars 64 forks source link

Patchfusion for relative depth to solve sky and ceiling problem #32

Closed andrewivan123 closed 2 months ago

andrewivan123 commented 6 months ago

Since metric depth estimation means that the depth of the sky will not be realistic, is it possible to train Patchfusion for relative depth estimation to solve the sky issue?

zhyever commented 4 months ago

Hi, some discussions can be found: https://github.com/zhyever/PatchFusion/issues/24

For me, the metric depth estimator provides absolute depth, which is somehow useful in certain practical cases. It would be better to investigate a possible solution for this issue just in a metric setting.

But anyway, it's not difficult to train PatchFusion for relative depth. We just need to change the loss function as far as I can see.