Averse Deep Semantic Segmentation

Abstract

Semantic segmentation consists in predicting whether any given pixel is part of the object of interest or not. Two types of errors are therefore possible: false positives and false negatives. For visualization and emphasis purposes, we might want to put special effort into reducing one type of error in detriment of the other. A common practice is to define the two types of errors as a relative trade-off using a cost matrix. However, it might be more natural for humans to define the trade-off in terms of an absolute constraint on one type of errors while trying to minimize the other. Previously, we suggested possible approaches to introduce this absolute trade-off in binary classifiers. Extending to semantic segmentation, we propose a threshold on the sigmoid layer and modifications to gradient descent such as adding a new term to the loss function and training in two phases. The latter produced the more resilient results, with a simple threshold being sufficient in most cases.