The preprocessed label files for KITTI-360 seems to miss the categories beyond 10

[npurson]

Upon sampling and analyzing several preprocessed label files, I discover that none of them contain classes greater than 10.

To reproduce this issue, you can follow the steps below in Python:

>>> path = 'kitti_360/data_2d_raw/2013_05_28_drive_0005_sync/voxels/000500.label'
>>> np.unique(np.fromfile(path, dtype=np.uint16).astype(np.float32))
array([ 0., 10., 20., 40., 44., 48., 49., 50., 52., 70., 72., 80., 99.],
      dtype=float32)
>>> path_label = 'kitti_360/labels/2013_05_28_drive_0005_sync/000500_1_1.npy'
>>> np.unique(np.load(path_label))
array([  0.,   1.,   5.,   6.,   7.,   8.,   9.,  10., 255.],
      dtype=float32)

[ahayler]

I encountered the same issue and after plotting I came to the conclusion that the NuScene label map was used to generate the KITTI360 preprocessed ground truth (this also maps a lot of classes to unlabeled). You need to rerun the label preprocessing with the KITTI360 label map (they can be found in the according .yaml files).

Hope that fixes your issue!

[Gaaaavin]

The label map for KITTI-360 is now uploaded. Please use the config file for pre-processing.