search

DepthAnything / Depth-Anything-V2

[NeurIPS 2024] Depth Anything V2. A More Capable Foundation Model for Monocular Depth Estimation
https://depth-anything-v2.github.io
Apache License 2.0
3.99k stars 345 forks source link

Metric model question #196

Closed ze-n-g closed 2 weeks ago

ze-n-g commented 4 weeks ago 
          Hi, I got a bad result when I used the metric/run.py and depth_anything_v2_metric_vkitti_vitl.pth to test the outdoor image.

image However, the result is right for using run.py. image The original image is here. ![Uploading image.png…]()

@duyuwen-duen, I have used your outdoor image, and I still got a bad result. image

Could someone provide some suggestions?

Originally posted by @szhang963 in https://github.com/DepthAnything/Depth-Anything-V2/issues/93#issuecomment-2252420260

LiheYoung commented 2 weeks ago

Hi, since the metric depth model is only trained on Virtual KITTI 2, it is not as robust as the relative depth models that are trained on large-scale data. Besides, for very far regions, e.g., sky, the metric depth is very challenging to estimate.