Action recognition from video

Let’s now learn how to use the MMAction toolbox (https://github.com/open-mmlab/mmaction) from the open-mmlab project to perform action recognition. The major features of MMAction are:

• Action recognition on trimmed videos (portion of the video that has an action)
• Temporal action detection (action localization) in untrimmed videos
• Spatial (parts of a frame that indicate an action) and temporal (variation of action across frames) action detection in untrimmed videos
• Support for various action datasets
• Support for multiple action understanding frameworks

First, let us understand how action recognition works. A video is a collection of images that are spaced over time (frames). We have two options of model input – 2D and 3D. 2D model input has a dimension of FxCHW where F is the number of frames and C, H, W are channels, height, and width respectively. 3D model input has an input dimension of CFHW...