I am currently working on a project that involves extracting features from LiDAR point cloud from the KITTI dataset. Despite exploring various resources and methods, I have not been able to find a solid/easy-to-implement solution. I am particularly interested in:

• Using pre-trained deep learning models for this task, such as PointNet, PointNet++, or other models available.

• Using mmdetection3d but other library/solution will also work. Although there is extract_feat function in mmdetection3d, I couldn’t find any documentation on how to implement that.

For example, to extract feature representation from the point cloud data (.bin files), I was expecting a code similar to this using openmmlab. However, I got an error as batch_inputs_dict['voxels'] does not exist. This indicates I may need to voxelize the data before passing it to the extract_feat function but I can’t find a work around for that.

from mmdet3d.apis import init_model, inference_detector

config_file = 'xyz.py'
checkpoint_file = 'xyz.pth'
point_cloud_path = 'abc.bin'

model = init_model(config_file, checkpoint_file, device='cuda:0')
result, data = inference_detector(model, point_cloud_path)
pcd_features = model.extract_feat(data)

I am reaching out to this community to seek any pointers to resources, code examples, or detailed explanations.

Thanks