Images/video have enormous pixel counts — feeding directly to fully connected (Dense) neural nets causes parameter explosion and fails to learn translation invariance. Convolutional Neural Networks (CNN) solve this by sliding small filters across the entire image to extract local patterns. MLS-C01 asks not CNN's internal equations but convolution/pooling roles, image classification vs object detection, SageMaker builtin/architecture choices.
A 224×224×3 color image flattened to a Dense layer has ~150k input nodes. First hidden layer of 1000 nodes = 150M weights — explosion. CNN solves this with three ideas.