Uncolored clouds and PointNet baseline #132
Conversation
There was a problem hiding this comment.
Hi @Vynikal, thanks a lot for your contribution!
It looks mostly ok to me, and I think it would be great to have this added to myria3d!
Before that, it needs some changes related to your readme and missing tests
| @@ -1,37 +1,60 @@ | |||
| <div align="center"> | |||
|
|
|||
| # Myria3D: Aerial Lidar HD Semantic Segmentation with Deep Learning | |||
| # Fork adapted to train/infer on non-colorized data | |||
There was a problem hiding this comment.
Can you remove the parts that are relate to your fork from this MR please, as it won't be relevant once merged
| [](https://github.com/IGNF/myria3d/actions/workflows/gh-pages.yml) | ||
| </div> | ||
| <br><br> | ||
| To train PointNet: |
| ___ | ||
|
|
||
| # Comparisons |
There was a problem hiding this comment.
I think that your comparisons can have a place in the documentation too, provided that you update the links so that the images are stored in our repo as well (in case yours is modified)
| @@ -0,0 +1,126 @@ | |||
| import torch | |||
There was a problem hiding this comment.
This needs testing, tests can be added in https://github.com/IGNF/myria3d/blob/main/tests/myria3d/models/modules/ (see https://github.com/IGNF/myria3d/blob/main/tests/myria3d/models/modules/test_randla_nets.py for an example)
| @@ -0,0 +1,46 @@ | |||
| # function to turn points loaded via pdal into a pyg Data object, with additional channels | |||
There was a problem hiding this comment.
this needs tests as well, that can be added in https://github.com/IGNF/myria3d/tree/main/tests/myria3d/pctl/transforms
Uncolored clouds
The motivation is to use the model as an augmentation of FLAIR dataset without using inaccurate cloud colorizing. As the attributes are different, it's necessary to write a new dataset description and pre-transform. Also rgb normalizing is now conditional to to rgb attribute. NoRGB separate experiments were created.
PointNet baseline
Taking advantage of the modular design of Myria3D, just another model to compare with, nothing more, nothing less. Also coming with the NoRGB variant.
The two NoRGB models were trained and tested, more info in the forked repository readme, also including trained checkpoints.
It might not be perfect, feel free to point out the issues.