DepthAnything / Depth-Anything-V2

Depth Anything V2. A More Capable Foundation Model for Monocular Depth Estimation
https://depth-anything-v2.github.io
Apache License 2.0
3.34k stars 270 forks source link

Questions about the teacher model #153

Open li12138cmd opened 2 weeks ago

li12138cmd commented 2 weeks ago

Hello, thank you very much for your work. I would like to ask, after using the teacher model to assign pseudo labels to the unlabeled data, will the images with prediction errors be filtered out? Or are they all directly fed into the network for learning?

LiheYoung commented 1 week ago

Hi @li12138cmd, we mentioned in our Section 5.2 that, for each pseudo-labeled sample, we ignore its top-n-largest-loss regions during training, where n is set as 10%. We consider them as potentially noisy pseudo labels.