How training depth anything v2?

AI Discussions
1

I’m confused about the intended workflow:

  1. Goal: Train a new model and release it as DepthAnythingV2.
  2. Current setup: The training script is initializing from checkpoints/DepthAnythingV2.ckpt.
  3. My understanding:
  • Normally, to train version 2, we’d start from the previous version’s weights (e.g. DepthAnythingV1.ckpt).
  • Then after training completes, we’d call the result DepthAnythingV2.
  1. Questions:
  • Why are we initializing from DepthAnythingV2.ckpt instead of DepthAnythingV1.ckpt?
  • Am I missing a step in the versioning or fine‑tuning process?
  • Is there a reason to continue training from an existing V2 checkpoint rather than from V1?

or maybe code to train DepthAnyhingV2 and V1 is totally private, and you just give the checkpoints to use in retraining with Metric_depth?

[Asked in this issue in github: Depth-Anything-V2 #291]

also I should add that when Im training never touch 0.9 delta1 and it’s just ranging around 0.8, while in paper it’s at least 0.936!

2

What course are you studying?