Questions about the settings in mmdet3d version

SamsungLabs / imvoxelnet

[WACV2022] ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection

MIT License

282 stars 29 forks source link

Questions about the settings in mmdet3d version #10

Closed gujiaqivadin closed 3 years ago

gujiaqivadin commented 3 years ago

Hello, saic-vul! Thanks for your excellent work imvoxelnet! I am looking into the realization in mmdet3d official repo of kitti dataset, and I am curious about anchor_generator settings in imvoxelnet_kitti-3d-car.py. I found that two anchor-generator are defined in this file, one for genrating anchors and the other for generating voxel-wise features. But these two settings are different in ranges(height dimensions). I want to know why it is different. Thanks a lot!

filaPro commented 3 years ago

Hi, @gujiaqivadin , Sure, these 2 anchor generators are the same for x and y axis and differs in z. The thing is Anchor3DHead is based on 2D convolutions, so Anchor3DHead.anchor_generator indeed returns 2D anchors. Which means that z_min = z_max and this 2 numbers are meaningless. However for voxel generation we utilize the range from -3 to +1 meters for z axis.

gujiaqivadin commented 3 years ago

OK, it solves my questions well. Thx!