Circular convolutions and infering back to 3D?

[JohanBergius]

Hi, Author. Could you please explain or give me directions to where I can better understand the approach of Circular convolutions (ring CNN). I would also like to understand better how your was able to reverse back to the original 3D space after going removing the Z dimension!

[YangZhang4065]

Hi Johan, sorry about the late reply.
For circular padding, please refer to my latest reply here: #46 (comment)
Reversing back to 3D space is actually quite easy. At the last layer of our neural network, we predict a 2D feature map of size HxWxZ for each scan/ BEV image. H and W is the height and width of the BEV map. Z is the feature dimension which equals to the number of semantic segmentation classes * the number of cells you want to have per each pillar (we used 32, but you can use whatever you want). Then you can reshape it to H x W x num_class x num_voxel_per_pillar and treat this reshaped feature map as semantic segmentation prediction on the 3D voxel level.