Network architecture inconsistent with paper about pointnet

[charlesq34]

Hi @shangyouai

Thanks for pointing that out! We found that T-net doesn't help much for this specific task so it's not included in the released code. The concatenation from 1024-dim feature seems result in slightly better performance -- it is slightly different from Fig. 2 as you pointed out. We shall correct that sentence if we would update the paper on ArXiv.

The T-net idea originates from Spatial Transformer Network (https://arxiv.org/abs/1506.02025), if you have known that it will become very natural. The benefits of using transformer network on point cloud is that the transformation computation is very simple and flexible -- it's just a matrix multiplication (in images or voxels you need to do bilinear or trilinear sampling).

Hope it helps. Let me know if you have more questions.

[charlesq34]

i'm closing the thread since no new comment comes.