Deep Learning for Computer Vision

Using 3D convolution for temporal learning

A video can be classified with 3D convolution. 3D convolution operation takes a volume as input and outputs the same, whereas a 2D convolution can take a 2D or volume output and outputs a 2D image. The difference is shown as follows:

Reproduced from Tran et al.

The first two images belong to 2D convolution. The output is always an image. 3D convolution, meanwhile, outputs a volume. The difference is a convolution operation in 3 directions with the kernel. Tran et al. (https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Tran_Learning_Spatiotemporal_Features_ICCV_2015_paper.pdf) used 3D ...