search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.24k stars 243 forks source link

Training with Synthetic / Dense Depth Annotations #116

Closed AlfieBrash closed 3 years ago

AlfieBrash commented 3 years ago

Instead of using sparse velodyne depth data, is it possible to implement training using dense depth annotations? (Grayscale 16bit PNG files obtained from synthetic data). Are there pre-made alternative options for depth_type for the config files? Thank you

VitorGuizilini-TRI commented 3 years ago

Which dataset class are you going to use? The depth_type key could be anything and depends on the dataset, on KITTI you can add another folder inside proj_depth with the name you want and just set depth_type to that. We have done this internally for VKITTI2, I'll look into adding support for that dataset here as well.

AlfieBrash commented 3 years ago

I am also working with vKITTI 2. I would use the raw png files (or jpg for loading?). Did you need to implement a grayscale-depth decoder in your internal work? Thank you very much for the help.

VitorGuizilini-TRI commented 3 years ago

Hi, sorry for the late reply. What do you mean by grayscale - depth decoder? For VKITTI2 you can still use RGB images as input, and convert the PNG depth maps to metric depth for supervision, then everything else remains the same.