Ques about details.

[huixiancheng]

Thx a lot for open source work! I have some questions listed below：

• Can polarmix be integrated into range-based approach? I tried it briefly on SalsaNext, but it didn't work well.
• Do it need more epochs for training after adding Polarmix? In my experiments, the convergence is slower and the final loss value is larger compared to the baseline.
• Is it necessary to rotate and paste Class==0 （'car'）? From the results, the performance improvement is not significant. However, considering the large number of Car points in each frame of the point cloud, replication enhancement of them may have potentially adverse effects.

  polarmix/core/datasets/semantic_kitti_polarmix.py

  Lines 51 to 63 in d0b7a24

  	kept_labels = [
  	'road', 'sidewalk', 'parking', 'other-ground', 'building', 'car', 'truck',
  	'bicycle', 'motorcycle', 'other-vehicle', 'vegetation', 'trunk', 'terrain',
  	'person', 'bicyclist', 'motorcyclist', 'fence', 'pole', 'traffic-sign'
  	]
  	"""
  	{'car': 0, 'bicycle': 1, 'motorcycle': 2, 'truck': 3, 'other-vehicle': 4, 'person': 5, 'bicyclist': 6, 'motorcyclist': 7,
  	'road': 8, 'parking': 9, 'sidewalk': 10, 'other-ground': 11, 'building': 12, 'fence': 13, 'vegetation': 14, 'trunk': 15,
  	'terrain': 16, 'pole': 17, 'traffic-sign': 18}
  	"""
  	instance_classes = [0, 1, 2, 3, 4, 5, 6, 7]
  	Omega = [np.random.random() * np.pi * 2 / 3, (np.random.random() + 1) * np.pi * 2 / 3] # x3