Mastering Computer Vision with TensorFlow 2.x

Spatial Pyramid Pooling

Most of the CNN models we introduced in Chapter 5, Neural Network Architecture and Models, require a fixed input image size, which limits the aspect ratio and scale of an input image. The fixed size constraint does not come from the convolution operation; instead, it comes from the fully connected layer, which requires a fixed input size. The convolution operation generates feature maps from the edges, corners, and different shapes of the images in different layers of the CNN. The feature maps are different in different layers and are a function of the shape within the image. They don't change considerably as a function of input size. SPP replaces the last pooling layer, just before the fully connected layer, and consists ...