updated lidar_hd_pre_transform function

[liubigli-tcp]

Hello!

Thank you very much for publishing this framework! It is really amazing!

I have created this pull request to allow any user to work with point clouds that have custom sets of features.
With this commit, the user specifies the list of point cloud features in its dataset_description yaml file.
Let assume that the point clouds in your dataset don't have any color information, the user can simply pass an empty
list in the color_keys and updates the number of input features (d_in) and that's all.

Please let me know if this could be of any interest to you :)

[leavauchier]

Hi @liubigli-tcp
Thanks a lot for your contribution!
It looks interesting to be more flexible on the input file (especially for people who have the possibility to train a new model).
I'll see with the rest of the team to have it intergrated soon

[leavauchier]

Hi @liubigli-tcp,
after a second thought, I don't think modifying lidar_hd_pre_transform is the right thing to do:
The architecture of the code was made so that you can create a pretransform function which is specific to your data (lidar_hd_pre_transform correspond to the pre-transformation for the LidarHD French lidar data program) and then tell myria3d which pre-transform method to use for your data in the configuration files here:

myria3d/configs/datamodule/hdf5_datamodule.yaml

Line 13 in 6e47ce3

- "${get_method:myria3d.pctl.points_pre_transform.lidar_hd.lidar_hd_pre_transform}"

More information on that in the documentation: https://ignf.github.io/myria3d/tutorials/prepare_dataset.html

However, creating a generic pre_transform could be possible (but without adding new arguments to the method that would be used as a points_pre_transform.

[liubigli-tcp]

Hi @leavauchier,

Thank you so much for your response! I really appreciate it. I’m currently working with point clouds that differ slightly from the LidarHD French lidar program, and I was wondering if there’s a way to adapt the code for other types of point clouds without having to write a new pre-transform function each time—just by adjusting the config files. This way, I would only need a simple system to manage my set of config files.

From what you’ve explained, it seems my approach might diverge too much from the way you originally designed the framework. For example, I would still need to add a custom function for each new type of point cloud.

If I also need to maintain the code for implementing my custom pre-transform functions, do you think it would make more sense for me to fork your repository and make these changes in my own version?