In PointNet++, suppose the input point data shape is [1, 3, 1024, 32] where the 1024 the centroids, and 32 is the neighbors and 3 is xyz coordinate with batch size as 1.
The input data is passed into nn.Conv2d(in_channels=3, out_channels=64, kernel_size=(1, 1))
Can we use nn.Linear instead of nn.Conv2d:
Change the input point data shape as [1 * 1024 * 32, 3], then pass the data into nn.Linear(3, 64)
I am wondering if these two have the same effect, since the conv2d kernel size is (1, 1)?