Asking for D2 Notation of session 3 in your paper

[seonhoon1002]

Hello, I raise this issue to asking for some notation of your paper. In your paper, you note that "From D3 to D0, the set of strides of their four stages for each model are {1,2,4,8}, {1,2,4,4}, {1,2,2,2} and {1,1,1,1}, respectively" So I understood D notations such as: D3: [1,2,4,8] D2: [1,2,4,4] D1: [1,2,2,2] D0: [1,1,1,1] But you note that "PointPillar model D2", and I understood PointPillar model has [1,2,2,2] stride(following your noation, it is D1). So is it notation miss or I misunderstood the pointPillar?

thanks

[Abyssaledge]

Thanks for your interest. Before further discussion, may I ask where you find the PointPillars of strides [1, 2, 2, 2] ?

[seonhoon1002]

I found it in PointPillar Paper and recheck in code in OpenPcdet pointpillar.yaml Like bellow:

"BACKBONE_2D: NAME: BaseBEVBackbone LAYER_NUMS: [3, 5, 5] LAYER_STRIDES: [2, 2, 2] NUM_FILTERS: [64, 128, 256] UPSAMPLE_STRIDES: [1, 2, 4] NUM_UPSAMPLE_FILTERS: [128, 128, 128] "

and mmdetection3D configuration too("hv_pointpillars_secfpn_kittti.py")

"model = dict( type='VoxelNet', voxel_layer=dict( max_num_points=32, # max_points_per_voxel point_cloud_range=[0, -39.68, -3, 69.12, 39.68, 1], voxel_size=voxel_size, max_voxels=(16000, 40000) # (training, testing) max_voxels ), voxel_encoder=dict( type='PillarFeatureNet', in_channels=4, feat_channels=[64], with_distance=False, voxel_size=voxel_size, point_cloud_range=[0, -39.68, -3, 69.12, 39.68, 1]), middle_encoder=dict( type='PointPillarsScatter', in_channels=64, output_shape=[496, 432]), backbone=dict( type='SECOND', in_channels=64, layer_nums=[3, 5, 5], layer_strides=[2, 2, 2], out_channels=[64, 128, 256]), "

[Abyssaledge]

The name layer_strides is a little bit misleading. I think layer_strides = [2, 2, 2] means downsampling feature maps 3 times with stride 2. So in terms of feature map sizes, the real network stride is [2, 4, 8]. And in Waymo Open Dataset, the PointPillars in MMDetection3D uses layer_strides=[1, 2, 2], where the feature maps only downsample 2 times. So in terms of feature map sizes, the real network stride is [1, 2, 4]. We split the last stage into two stages, so in our notations, the stride of D1 is [1, 2, 4, 4]. I hope I made it clear. Let me know if you have further questions.

[seonhoon1002]

We may have the different view of "stride" term.

Thanks for nice anwser!