V2AI / Det3D

World's first general purpose 3D object detection codebse.

https://arxiv.org/abs/1908.09492

Apache License 2.0

1.48k stars 299 forks source link

Wrong computation of f_center in PillarFeatureNet #92

Closed zwqnju closed 4 years ago

zwqnju commented 4 years ago

In https://github.com/poodarchu/Det3D/blob/b4cb6428b807d8cfe2135b535961d710ff9d41c1/det3d/models/readers/pillar_encoder.py#L129 it is to compute the x distance between each point and its corresponding pillar's center. But the current code does not use the point's x value. This line should be modified to: f_center[:, :, 0] = features[:, :, 0] - ( The computation in y axis has the similar bug, and should be modified too: https://github.com/poodarchu/Det3D/blob/b4cb6428b807d8cfe2135b535961d710ff9d41c1/det3d/models/readers/pillar_encoder.py#L132

68 also mentioned this issue.

tianweiy commented 4 years ago

@zwqnju this is a bug. Besides, here is also incorrect for pointpillar on nuScene. Basically, the timestamp for different sweeps is discarded

zwqnju commented 4 years ago

@tianweiy Oh, I see. Thanks. I also found another bug: In the construction method of PillarFeatureNet, voxel_size and pc_range both have default values. In the config file like this, the voxel_size and pc_range are different with the default values, but the config file do not set them explicitly in model.reader. This also results in the wrong computation of f_center.

poodarchu commented 4 years ago

you can make a mr.