Pyramid Pooling implemented in PyTorch

This Module implements Spatial Pyramid Pooling (SPP) and Temporal Pyramid Pooling (TPP) as described in different papers.

Temporal Pyramid Pooling:

Sudholt, Fink: Evaluating Word String Embeddings and LossFunctions for CNN-based Word Spotting

Principle

Given an 2D input Tensor, Temporal Pyramid Pooling divides the input in x stripes which extend through the height of the image and width of roughly (input_width / x). These stripes are then each pooled with max- or avg-pooling to calculate the output.

Animated Principle

Spatial Pyramid Pooling:

He, et. al.: Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

Principle

Given an 2D input Tensor, Spatial Pyramid Pooling divides the input in x² rectangles with height of roughly (input_height / x) and width of roughly (input_width / x). These rectangles are then each pooled with max- or avg-pooling to calculate the output.