Deep Learning - 3D Convolutional Neural Network (3DCNN)

1. 3DCNN understands that
2D convolution only considers the spatial information of 2D pictures, so it is only suitable for visual understanding tasks of a single 2D picture. When processing 3D images or videos, the input of the network has one more dimension, and the input consists of $(c,height,width)$变为了$(c,depth,height,width)$ , where$c$ is the number of channels,$depth$ is the width of the input data. Therefore, when processing the data, the convolution needs to be transformed accordingly, from 2D convolution to 3D convolution.
On the basis of 2D convolution, 3D convolution is proposed. 3D convolution has one more dimension than 2D convolution in structure, and the size of 2D convolution can be expressed as$k_h\times k_w$, and the size of the 3D convolution can be expressed as $k_h\times k_w\times k_d$. The specific calculation formula of 3D convolution is similar to that of 2D convolution, that is, each time you slide, it is the same as $c$ channels, the size is$(depth,height,width)$ is multiplied and added to obtain a value in the output feature map, as shown in the figure.

视频输入的维度：$inpu{t}_C\times inpu{t}_T\times inpu{t}_W\times inpu{t}_H$;
Dimensions of 3D convolution kernel: $inpu{t}_C$A parallel dimension is $kerne{l}_T\times kerne{l}_W\times kerne{l}_H$The convolution kernel;
3D convolution kernel in $T,W,H$ moves in three directions.

Reference:
3D CNN