Figure 5.

The comparison Figure of 2D convolution and 3D convolution. (a) 2D convolution diagram: the 2D convolution kernel convolves with a single image to obtain a 2D feature map. (b) 3D convolution diagram: the input is a 3D cube composed of multiple consecutive video frames that can be expanded into multiple 2D images in temporal series. The size of the 3D convolution kernel in the temporal dimension is 3. The 3D convolution kernel convolves with multiple consecutive video frames to obtain multiple feature maps. The connecting lines of shared weights are in the same color. Two different 3D convolution kernels can extract two types of features and generate two sets of different feature maps on the right.