Scale-Aware Cascading for Semantic Segmentation

Abstract

Semantic segmentation is an important technology commonly used in medical imaging, autonomous driving vehicles, and backgrounds for virtual meetings. Scale Aware approaches have become the standard when it comes to the semantic segmentation domain of Machine Learning. Multiple image scales are passed through the network allowing the result to use the regular CNN layers such as max-pooling as well as convolution layers. Also, a cascading hierarchy of attention has been shown to improve the accuracy of models for such segmentation tasks. The combination of both these approaches has been shown to greatly improve the accuracy of such models. A side effect of using the cascading approach is that the model turns out to use less memory in comparison to previous approaches. Auto-labelling engines are also helpful in generalizing the model further. The cityscapes dataset used here is a useful data bank as it consists of a myriad of situations where the model can be trained and tested on. This paper presents the tested results of such a segmentation model and incremental modifications to the model pipeline to understand and improve upon the existing architecture.