[ICCV 2021] HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration

Thank for your generous sharing. However after I reproduced your experiment, I found that the results of the rotation error in the KITTI_SUMMARY is not consistent with the results in the paper. How did you get it from the results in the KITTI_SUMMARY?

I think there's a mistake in Hierarchical Transformation from coarse to fine, the last element in transformation matrix should be 1 not 0, it will result the wrong translation. Maybe it can be fixed like this:

I ran the code sh scripts/train_kitti_det.sh, and I got error message like this:

I looked into it on the google, they said it was because batch_size is too large or num_worker is too large.

I set parameter "batch_size" as 1, and set parameter "gpu" as "0,1,2"(everyone is a 3090 gpu), and set train_feats.py/train_loader/num_workers as 1，but I still got the same message.

Here is my train_kitti_det.sh by the way.

python train_feats.py --batch_size 1 --epochs 100 --lr 0.001 --seed 1 --gpu 0,1,2 \
--npoints 16384 --dataset kitti --voxel_size 0.3 --ckpt_dir /media/data3/hlf_data/HRegNet0/HRegNet/ckpt \
--use_fps --use_weights --data_list ./data/kitti_list --runname "train_kitti_det0" --augment 0.5 \
--root /media/data3/hlf_data/HRegNet0/HRegNet/data/kitti_list --wandb_dir /media/data3/hlf_data/HRegNet0/HRegNet/wandb_env --use_wandb

Is there any way to solve this? :)

I run the command sh scripts/train_kitti_det.sh, and I got the error message like this:

Traceback (most recent call last):
  File "/media/data3/hlf_data/HRegNet0/HRegNet/train_feats.py", line 243, in <module>
    train_feats(args)
  File "/media/data3/hlf_data/HRegNet0/HRegNet/train_feats.py", line 159, in train_feats
    for i, data in pbar:
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/tqdm/std.py", line 1195, in __iter__
    for obj in iterable:
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
    data = self._next_data()
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1085, in _next_data
    return self._process_data(data)
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 1111, in _process_data
    data.reraise()
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/_utils.py", line 428, in reraise
    raise self.exc_type(msg)
FileNotFoundError: Caught FileNotFoundError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/utils/data/_utils/worker.py", line 198, in _worker_loop
    data = fetcher.fetch(index)
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/torch/utils/data/_utils/fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/media/data3/hlf_data/HRegNet0/HRegNet/data/kitti_data.py", line 89, in __getitem__
    src_points = read_kitti_bin_voxel(data_dict['points1'], self.npoints, self.voxel_size)
  File "/media/data3/hlf_data/HRegNet0/HRegNet/data/kitti_data.py", line 20, in read_kitti_bin_voxel
    scan = np.fromfile(filename, dtype=np.float32, count=-1).reshape([-1,4])
FileNotFoundError: [Errno 2] No such file or directory: '/media/data3/hlf_data/HRegNet0/HRegNet/data/kitti_list/01/velodyne/000204.bin'

Seems like I didn't process the data files rightly.

I see it in the readme.md that the files should be organized like this:

But I don't know how to come to this. Mine is just like thiis:

Someone help me please? Thanks!

I was running sh scripts/train_kitti_det.sh,then I got the error message

/home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/MinkowskiEngine-0.5.4-py3.9-linux-x86_64.egg/MinkowskiEngine/__init__.py:36: UserWarning: The environment variable `OMP_NUM_THREADS` not set. MinkowskiEngine will automatically set `OMP_NUM_THREADS=16`. If you want to set `OMP_NUM_THREADS` manually, please export it on the command line before running a python script. e.g. `export OMP_NUM_THREADS=12; python your_program.py`. It is recommended to set it below 24.
  warnings.warn(
Traceback (most recent call last):
  File "/media/data3/hlf_data/HRegNet0/HRegNet/train_feats.py", line 7, in <module>
    from data.kitti_data import KittiDataset
  File "/media/data3/hlf_data/HRegNet0/HRegNet/data/kitti_data.py", line 11, in <module>
    from models.utils import generate_rand_rotm, generate_rand_trans, apply_transform
  File "/media/data3/hlf_data/HRegNet0/HRegNet/models/utils.py", line 7, in <module>
    import point_utils_cuda
ImportError: /home/hlf/miniconda3/envs/HRegNet/lib/python3.9/site-packages/point_utils-0.0.0-py3.9-linux-x86_64.egg/point_utils_cuda.cpython-39-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda20getCurrentCUDAStreamEa

I checked my file list, found no file named "point_utils_cuda.py" indeed.

How can I deal with this problem? Thank you very much!

Hello,

Thank you for the sharing but I found out if I transfer your datalist (e.g. kitti sequence 08) back to the format in kitti odometry, they don't equal to each other. can you tell me how did you process to get your train data set ?

Hello,

Thank you for your generosity to share the code. However I got stuck when I come to the forward() function of CoarseReg class.

What I understand about line 249 and line 255 is that you want to gather k nearest cosine neighbors in src for keypoints in dst and k nearest neighbors in dst for keypoints in src. I didn't understand why you used src_knn_idx for both cases? I assume it should be dst_knn_idx instead in line 249, because it gathers k nearest neighbors in src to keypoints in dst.

Best regards.

Hello, I am trying to use just the keypoint detector part and am willing to use the pretrained detector weights to be tested on my own data sample (save keypoints of each pointcloud). I can see that there is a test.py for the testing entire network, could you please guide me on how can I test just the detector part and not the entire Net on my sample?

Thanks for the nice work, btw. Looking forward to your response. Fatemeh

Hello,

I have been studying your model and I found out that the keypoints are found by summing the product of weights and knn points, how ever this may not affirm that they exactly exist in the original point cloud. Besides I found something silimar in the process of finding correspondances in target point cloud that you use the attentive weights to multiplys knn points as well, which doesn't verify the corres points exist in target point cloud either. Did I get somewhere wrong ?