search

darglein / ADOP

MIT License
2.02k stars 197 forks source link

multi-orientation scene #65

Closed klae01 closed 2 years ago

klae01 commented 2 years ago

I'm interested in your repositories and are working on training a model.

However, I found out that there is a bug that cannot be learned when shooting with the camera upright. (in exif term, Orientation |Right-top) The size of tensor a (3024) must match the size of tensor b (4032) at non-singleton dimension 3

So, I suggest some methods. swap w and h in camera.ini according to orientation.

But, the results do not seem to improve over time. it make only noisy result.

any other specific parameter do i have to change? I try to solve this orientation issue

klae01 commented 2 years ago

solved by reorient image and modify exif data.

  1. convert image data by https://github.com/klae01/image-reorientation
  2. colmap reconstruction
  3. ADOP train

You can now train successfully.