Towards Semantic Segmentation Using Ratio Unpooling

This paper presents the concept of Ratio Unpooling as a means of improving the performance of an Encoder-Decoder Convolutional Neural Network (CNN) when applied to Semantic Segmentation. Ratio Unpooling allows for 4 times the amount of positional information to be carried through the network resulting in more precise border definition and more resilient handling of unusual conditions such as heavy shadows when compared to Switch Unpooling. Applied here as a proof-of-concept to a simple implementation of SegNet which has been retrained on a cropped and resized version of the CityScapes Dataset, Ratio Unpooling increases the Mean Intersection over Union (IoU) performance by around 5–6% on both the KITTI and modified Cityscapes datasets, a greater gain than by applying Monte Carlo Dropout at a fraction of the cost.