Feature extraction on point cloud using the pre-trained network

[shahriar0651]

Hello, I want to extract feature representation from the point cloud data (.bin files).

I was expecting a code similar to this, but 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.

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)

Is there any end-to-end documentation on how to run feature extraction on point cloud data, or do you have any idea how exactly I can fix it?

Thanks