Open Reed-yang opened 4 days ago
I think the comment reply in the issue: https://github.com/KTH-RPL/DeFlow/issues/3#issuecomment-2128533742 it explains that for evaluation you always need ground truth to do that. While for inference proposal, you don't need ground truth.
The only step you need to do is (write your dataset class to read data) or (transfer your data to our h5py file format):
group.create_dataset('lidar', data=pc.astype(np.float32))
group.create_dataset('ground_mask', data=gm.astype(bool))
group.create_dataset('pose', data=pose.astype(np.float32))
if flow_0to1 is not None:
# ground truth flow information
group.create_dataset('flow', data=flow_0to1.astype(np.float32))
group.create_dataset('flow_is_valid', data=flow_valid.astype(bool))
group.create_dataset('flow_category_indices', data=flow_category.astype(np.uint8))
group.create_dataset('ego_motion', data=ego_motion.astype(np.float32))
There are two examples how you can have a h5py file right now, check: https://github.com/KTH-RPL/SeFlow/tree/main/dataprocess
btw, is it possible to test on dataset that has no valid ground_mask
available? Or is there a convenient way to compute ground_height.npy
& se2.json
on my aggregated point cloud (from all frames)?
thx in advance, :)
You can directly try this python package to get ground_mask: https://github.com/Kin-Zhang/linefit/tree/master
One more question, I noticed that in process_log
, pose
and car_frame_pc
are input for model inference. What coordinate system is car_frame_pc
established on, Can you explain it in detail? In each frame, is car_frame_pc
ego-centric or is it based on a unified first moment as the origin (world coor)? What should I do if my data is in world coor?
I think it's fine for both, pc
is ego-centric and you will save pose for the world coor, and the first pose can be the original world coor. Here is an example dataloader I wrote one year ago which data is in world coor:
I saw issue on Deflow that eval on custum dateset needs groud truth flow. I'm very exciting that you have this self-supervised version of this model and wondering whether it is possible to inference scene flow using your pretrained weight on dataset: has frame point clouds, poses, and segmentation maybe, then producing flow results.