Scaling problem of metric depth when inferring on the NYUv2 dataset

[YkiWu]

Thanks for your work! When I was using the trained indoor metric depth model you provided to infer the depth on the NYUv2 dataset, I set the max_depth to the 20 you provided. The resulting point cloud seems to have some scaling differences from the real-world coordinates. Perhaps you know what I should set the max_depth to for the NYUv2 dataset? Or should I retrain a specific depth estimation model for the NYUv2 dataset?

[YkiWu]

By the way, the pretrained model I used is depth_anything_v2_metric_hypersim_vitl.pth

[LiheYoung]

Hi, the model here is trained on Hypersim. If you want to predict metric depth for NYUv2, you may need our metric depth models trained in Depth Anything V1, which is trained on NYUv2: https://github.com/LiheYoung/Depth-Anything/tree/main/metric_depth