Anchor3DRangeGenerator

[holtvogt]

I'm wondering why the z values within ranges differ from the point_cloud_range.

• Why -0.6 for pedestrians/cyclists and -1.78 for cars?
• Do I have to choose these values for my custom (KITTI-based) dataset as well?

Seen, e.g. here:

mmdetection3d/configs/_base_/models/hv_pointpillars_secfpn_kitti.py

Line 41 in 073f353

[0, -39.68, -0.6, 69.12, 39.68, -0.6],

[ZCMax]

Please refer to this issue:#986, you need to adjust these value for you own custom dataset.

[holtvogt]

So, as far as I understand, the z values should match the z position of my ground area in my recordings?

[Tai-Wang]

Should match the average height of object centers.

[wittmstSICKAG]

Since I am currently training SECOND and PointPillars on a custom dataset, I am also curious how to choose/adjust the settings for the Anchor3DRangeGenerator.

@Tai-Wang You mentioned that the z-coordinate should match the average height of object centers. Now, the Velodyne sensor in the KITTI setup is positioned at z = 1.73m (see setup). If this coordinate system is used for training, this would mean that the object centers relative to the ground are:

• Pedestrians / Bicycles: z = 1.73m + (-0.6m) = 1.13m
• Cars: z = 1.73m + (-1.78m) = -0.05m

Furthermore, if we now compute the bottom center of these bounding boxes by taking the anchor box dimensions into account, I end up with:

• Pedestrians / Bicycles: z = 1.13m - 1.73m/2 = 0.265m
• Cars: z = -0.05m - 1.56m/2 = -0.83m

Is this correct? To me it seems that the anchor z ranges for pedestrians and bicycles are correct but off for cars. If the object center is used, I would expect for cars a z offset of (1.56m/2) - 1.73m = -0.95m and not -1.78m.