Can someone explain how to use Conv3d and ConvND in caffe?

Question

Can someone please, explain how one can use a Conv3D or a ConvND for Depth-images or videos or pretty much any 3d (n-d?) data in Caffe ?
Is there any example or demo for Conv3D ?

Answer 1

You can use the regular "Convolution" layer to process blobs of any dimension. You only need to pay close attention to the parameters:

layer {
  type: "Convolution"
  name: "conv_nd"
  bottom: "in" # 5D blob 
  too: "out"
  convolution_param {
     kernel_size: 3
     kernel_size: 5
     kernel_size: 5 # define 3 by 5 by 5 kernel

     pad: 1
     pad: 2
     pad: 2  # pad according to kernel size

     stride: 1
     stride: 2
     stride: 2 # you can have different stride for different dimensions

     axis: 1  # the "channel" dimension
     num_output: 30 # output 30 dim per 3D voxel
  }
}

For more information read the comments on Convolution param in the caffe.proto file.