PointNet implementation

Hi, a PointNet-based network that processes each point in the masked 3D point cloud to a geometric feature embedding, said in the paper. But in the code implementation only convolution and ReLU of point cloud x in network.py. No spacial transform networks(STN) and maxpooling operation on point cloud. These are the two most important features of PointNet.

` Class PoseNetFeat (nn. Module): def init(self, num_points): super(PoseNetFeat, self).init() self.conv1 = torch.nn.Conv1d(3, 64, 1) self.conv2 = torch.nn.Conv1d(64, 128, 1)

    self.e_conv1 = torch.nn.Conv1d(32, 64, 1)
    self.e_conv2 = torch.nn.Conv1d(64, 128, 1)

    self.conv5 = torch.nn.Conv1d(256, 512, 1)
    self.conv6 = torch.nn.Conv1d(512, 1024, 1)

    self.ap1 = torch.nn.AvgPool1d(num_points)
    self.num_points = num_points
def forward(self, x, emb):
    x = F.relu(self.conv1(x))
    emb = F.relu(self.e_conv1(emb))
    pointfeat_1 = torch.cat((x, emb), dim=1)

    x = F.relu(self.conv2(x))
    emb = F.relu(self.e_conv2(emb))
    pointfeat_2 = torch.cat((x, emb), dim=1)

    x = F.relu(self.conv5(pointfeat_2))
    x = F.relu(self.conv6(x))

    ap_x = self.ap1(x)

    ap_x = ap_x.view(-1, 1024, 1).repeat(1, 1, self.num_points)
    return torch.cat([pointfeat_1, pointfeat_2, ap_x], 1) #128 + 256 + 1024`

I am confused. Would you explain this to me? Thanks in advance.

[j96w]

Hi, (1) Since we are already uncovering the 6DoF pose (rotation and translation) of the target object, we don’t need the STN to rotate the input pointcloud which might disturb the pose estimation process. (2) Maxpooling will lead to poor training efficiency of the CNN-based encoder-decoder network since it’s skipping too many pixel embeddings during the Max-pool step.

[roywithfiringblade]

though STN and symmetry function like max pooling are the key idea of PointNet, cause they are the operation to deal with the problem of pointcloud data , otherwise the whole architecture is actually not so “PointNet” …