What is 3D convolution and how is it used in video understanding models?

3D convolution extends 2D convolution by adding temporal dimension for video data.

3D convolution operates over width, height, and time, allowing models to learn motion features directly from video clips. Unlike 2D CNNs which process frames independently, 3D CNNs capture temporal dynamics like movement, action, and motion continuity, making them ideal for video classification and action recognition.