search

darglein / ADOP

MIT License
2.03k stars 199 forks source link

How to train with your own data? #38

Closed AdamMarciniak closed 2 years ago

AdamMarciniak commented 2 years ago

What are the inputs for training? Say I go outside and want to do the same for new places or buldings/objects. Do I need to take a series of images or a video? How do I provide or get the initial pointcloud data?

Also, is there a way to train this on a GPU without the memory that you use? I've got a GTX 1070 Ti with like 8 GB of ram. Is there a setting I can use to make that work?

darglein commented 2 years ago

The basic pipeline would be:

  1. Capture images
  2. Reconstruct with COLMAP (sparse + dense)
  3. Convert the colmap scene to adop
  4. Train the system

8GB of video memory is quite low but if you reduce the batch size it should still work (quality might be a bit worse tho).

yohannes-taye commented 2 years ago

How to convert a colmap scene to adop? @darglein

HandsomeFYM commented 2 years ago

How to convert a colmap scene to adop? @darglein

Maybe you will find the answer in this document. The link: https://github.com/darglein/ADOP/blob/main/scenes/README.md