search

nianticlabs / monodepth2

[ICCV 2019] Monocular depth estimation from a single image
Other
4.16k stars 963 forks source link

How to use depth maps instead of point clouds #446

Closed Shawn9257 closed 2 years ago

Shawn9257 commented 2 years ago

Thanks for your work. I am trying to use a dataset called MKV dataset(KinectDatasets) to train mono depth prediction. I am now trying to make the custom dataset as same as Kitti. I only have color images and depth images, not point clouds. Should I use some way to change depth images to point clouds? 0_00000000

daniyar-niantic commented 2 years ago

Hi @Shawn9257

From what I understand, you want to use Kinect depth data to train a network. I don't think this is the right repository for this task, as monodepth2 uses self-supervision (no depth data) and you need supervised monocular depth estimation.

I suggest to look into methods that train on NYU v2 Depth dataset.

Shawn9257 commented 2 years ago

Thanks for your reply! That really helped me a lot. The reason why I try to do this is that when I try to use the pre-train model to fine-tune with the Kinect depth dataset(without depth maps). It gets a bad result. The model does not seem to converge, and the depth prediction, which could be partially effective, is worse after training. Do you have any suggestions?