What is 3D CNN and where are they used (I3D, SlowFast)?

3D CNN (three-dimensional convolutional neural networks) is a type of convolutional network that uses 3D convolutions, meaning filters with three dimensions (height, width, and depth/time). Unlike standard 2D CNNs, which work with images, 3D CNNs process video data or sequences of images, taking into account spatiotemporal information.

Main applications of 3D CNN:

• Video analysis (action and event recognition)
• Medical imaging (3D scans, tomography)
• Processing image sequences

I3D (Inflated 3D ConvNet) is an architecture that extends 2D CNN into 3D by "inflating" 2D filters into 3D, allowing the use of pre-trained 2D weights and efficiently training models for video analysis.

SlowFast is an architecture that uses two parallel streams: a "Slow" stream with a low frame rate to capture spatial information and a "Fast" stream with a high frame rate to capture motion dynamics. This enables effective modeling of temporal aspects of video.

Both architectures are widely used in video analytics tasks, such as action recognition, video classification, and behavior analysis.