About depth ckpt

[frkmac3]

I fine-tuned UniDepth using THEIR official code on my own dataset. However, when comparing the keys of the UniDepth weights provided by DetAny3d with the keys from fine-tuned checkpoint, they do not fully match(even the total number of keys differs).
Does this mean det any 3d has made structural modifications to UniDepth?

[jjxjiaxue]

Hi—please see how we load it here: https://github.com/OpenDriveLab/DetAny3D/blob/main/train.py#L518C17-L527C53

We use the UniDepth v2 checkpoint, and some key names may differ.

[frkmac3]

Hi—please see how we load it here: https://github.com/OpenDriveLab/DetAny3D/blob/main/train.py#L518C17-L527C53

We use the UniDepth v2 checkpoint, and some key names may differ.

Yes, I've noticed that as well. Your depth model has 768 state dict keys, while the official UniDepthV2 checkpoint has only 520.

[jjxjiaxue]

I’m not sure why the counts differ — to my knowledge we did not modify UniDepth. The UniDepth checkpoint in our Google Drive is exactly the one downloaded from the official UniDepth repo. qaq

[frkmac3]

I’m not sure why the counts differ — to my knowledge we did not modify UniDepth. The UniDepth checkpoint in our Google Drive is exactly the one downloaded from the official UniDepth repo. qaq

I fine-tuned the model using my own dataset, but the results are quite poor. I suspect it's due to an inappropriate training configuration. Could you please release the fine-tuning code so that we can reference it?

[jjxjiaxue]

Sure — could you share your current fine-tuning configuration first? That will help me understand what might be causing the poor results and see whether it’s a config issue.

[jjxjiaxue]

Also, just a heads up — if your config specifies a finetuned depth ckpt and you resume from our pretrained ckpt, the training script will skip loading the depth weights. This may affect your results.