search

isl-org / ZoeDepth

Metric depth estimation from a single image
MIT License
2.24k stars 207 forks source link

Any suggestion about fine-tuning on another outdoor dataset #27

Open Gaozihui opened 1 year ago

Gaozihui commented 1 year ago

Thanks for your brilliant work.

I am trying to use the pre-trained model to fine-tune on another outdoor dataset. Should I use the pretrained model ZoeD_M12_K.pt or the pretrained model without fine-tuning on any dataset (which is not released I think) ? I have tried the former one but I don't see an obvious droping on the training loss. image

Thank you in advance for any advice you can offer.

MACILLAS commented 1 year ago

Hi @Gaozihui I am also trying to fine-tune on another outdoor dataset. I'm getting stuck thou when I run the train_mono.py with ZoeD_MK12_K.pt I get a bunch of size mismatch errors. I'm wondering if you had this issue and if you had any pointers?

Thanks in advance! 

python train_mono.py -m zoedepth -d kitti --pretrained_resource="url::https://github.com/isl-org/ZoeDepth/releases/download/v1.0/ZoeD_M12_K.pt"
Gaozihui commented 1 year ago

Hi @Gaozihui I am also trying to fine-tune on another outdoor dataset. I'm getting stuck thou when I run the train_mono.py with ZoeD_MK12_K.pt I get a bunch of size mismatch errors. I'm wondering if you had this issue and if you had any pointers?

Thanks in advance!

python train_mono.py -m zoedepth -d kitti --pretrained_resource="url::https://github.com/isl-org/ZoeDepth/releases/download/v1.0/ZoeD_M12_K.pt"

Hi~ Have you solved this problem? 

python train_mono.py -m zoedepth -d nuscenes --batch_size 8 \
 --trainer zoedepth --pretrained_resource "" config_version kitti

I think maybe you left out the 'config_version' term and the pretrained_source should be "". Please try it.

ariqshadi commented 1 year ago

Hi, im just wondering, when you fine tune using a custom dataset, are the label depth images for the training in a relative depth value or in a metric depth value? because my label images are in metric depth value and the result seems to be far off

thejerrycheng commented 6 months ago

I have the same problem here, should the custom dataset be metric depth or in relative depth

MACILLAS commented 6 months ago

I think the main idea is that the decoder learns metric depth, so you fine-tune with metric depth dataset. If you have a relative depth dataset I think you can just evaluate MiDASv3 on it directly?