karoly-hars / DE_resnet_unet_hyb

Depth estimation from RGB images using fully convolutional neural networks.
https://link.springer.com/chapter/10.1007%2F978-3-319-98678-4_38
BSD 3-Clause "New" or "Revised" License
55 stars 14 forks source link

Generate point cloud #5

Closed TuskAW closed 4 years ago

TuskAW commented 4 years ago

I am getting good results testing on my image set. I am wondering how I can generate a point cloud from the depth displayed with show_img_and_pred. Any advice will be much appreciated.

karoly-hars commented 4 years ago

I am glad you are getting good results.

Creating a point-cloud from a depth estimation is not a trivial task. I attempted to do something like this in one of my other repos: https://github.com/karoly-hars/3D_scene_reconstruction_CNNs. The results are not bad but also not that great and in order to make it work you need to know some of the camera params.

If you are trying to create a 3D environment from a depth estimation, using stixels (https://www.google.com/search?q=stixels&source=lnms&tbm=isch&sa=X&ved=2ahUKEwj4tuel0NTqAhXWvosKHQ8nBvYQ_AUoAXoECAwQAw&biw=1853&bih=949) is probably a better approach.

TuskAW commented 4 years ago

Thank you very much for the information! This gives me a good starting point for what I am trying to do.