locally aggregated feature and mask

[Wei Jiang]

Hi:
Thanks for sharing this project!
I'm trying to understand one function:

VoxelNet-pytorch/voxelnet.py

Lines 62 to 81 in bfd7be1

 class VFE(nn.Module): 
  
     def __init__(self,cin,cout): 
         super(VFE, self).__init__() 
         assert cout % 2 == 0 
         self.units = cout // 2 
         self.fcn = FCN(cin,self.units) 
  
     def forward(self, x, mask): 
         # point-wise feauture 
         pwf = self.fcn(x) 
         #locally aggregated feature 
         laf = torch.max(pwf,1)[0].unsqueeze(1).repeat(1,cfg.T,1) 
         # point-wise concat feature 
         pwcf = torch.cat((pwf,laf),dim=2) 
         # apply mask 
         mask = mask.unsqueeze(2).repeat(1, 1, self.units * 2) 
         pwcf = pwcf * mask.float() 
  
         return pwcf 

Especially:

VoxelNet-pytorch/voxelnet.py

Lines 73 to 74 in bfd7be1

 #locally aggregated feature 
 laf = torch.max(pwf,1)[0].unsqueeze(1).repeat(1,cfg.T,1) 

The mask is not used for generating the locally aggregated feature. This may cause the local feature is consisted of (0,0,0) points.

For example, if an input is in shape (B, C, N), where B is the batch size, C is the number of channel, N is the number of the points. And the first half part of N are all negative values, and the second half is empty, which represented as (0,0,0). The locally aggregated feature which in shape (B, C, 1), after max operation should be all zeros, which is wrong I think.

Bests,