Table 3.

3D-CNN network architecture.

Type/stride	Padding	Filter size	Number of filters
Conv/s1	Same	$3\times 3\times 1$	64
Max pool	Same	$4\times 4\times 2$	–
Conv/s1	Same	$3\times 3\times 1$	64
Conv/s1	Same	$3\times 3\times 1$	64
Conv/s1	Same	$3\times 3\times 1$	64
Conv/s1	Same	$3\times 3\times 1$	64
Conv/s1	Same	$3\times 3\times 1$	128
Max pool	Same	$2\times 2\times 1$	–
FC-256
FC-128

The input of the 3D-CNN is 3-D histograms $H\in {\mathbb{R}}^{T\times b\times d}$ and the 3-D convolution operation is performed over the ‘time’, ‘bin’, and ‘band’ dimensions.