search

LiheYoung / Depth-Anything

[CVPR 2024] Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data. Foundation Model for Monocular Depth Estimation
https://depth-anything.github.io
Apache License 2.0
6.7k stars 515 forks source link

Rgb-depth pairs for fine-tuning #56

Open abhishekmonogram opened 7 months ago

abhishekmonogram commented 7 months ago

Hi @LiheYoung, I wanted to finetune the metric depth estimation code on my own dataset. I wanted to ask you about the number of pairs of rgb-depth required to get good metrics on metric depth estimation? The NYUV2 dataset have ~36000 pairs of rgb-depth images in the training dataset. Do you think a few 100 or 1000 pairs maybe enough for accurate metric depth estimation?

cosmosmosco commented 7 months ago

Hello, may I ask which weight model you use for fine-tuning? I used depth_anything_vitl14.pth according to the ReadMe, but there was a mismatch error in state_dict. Then I tried depth_anything_metric_depth_outdoor.pt, and there was no such error. I wonder if you did the same?

oconnor127 commented 5 months ago

@abhishekmonogram I also want to fine-tune the model on a custom dataset. How did you do that?