Optimizing the Relevance-Redundancy Tradeoff for Efficient Semantic Segmentation

Semantic segmentation aims at jointly computing a segmentation and a semantic labeling of the image plane. The main ingredient is an efficient feature selection strategy. In this work we perform a systematic information-theoretic evaluation of existing features in order to address the question which and how many features are appropriate for an efficient semantic segmentation. To this end, we discuss the tradeoff between relevance and redundancy and present an information-theoretic feature evaluation strategy. Subsequently, we perform a systematic experimental validation which shows that the proposed feature selection strategy provides state-of-the-art semantic segmentations on five semantic segmentation datasets at significantly reduced runtimes. Moreover, it provides a systematic overview of which features are the most relevant for various benchmarks.