A Review on Semantic Segmentation from a Modern Perspective

摘要

Among various vibrant areas of research in computer vision, scene understanding is the one, which has garnered exceptional attention from the research community. This increased popularity should be credited to the accelerated research in deep learning techniques, mainly during the last decade. In scene understanding, one important step is semantic segmentation where the goal is to give each pixel of an image a proper object category label so that each object gets properly delineated. This technique is also known as scene parsing. One major challenge in this task is that it combines the problem of object detection, image segmentation and multi-class recognition. It encompasses a wide variety of practical applications such as autonomous driving, indoor navigation, virtual/augmented reality etc. This paper gives a review on semantic segmentation from a modern perspective by giving a special attention to deep learning based scene parsing methods. In this review, we take up two central issues of semantic segmentation-accuracy (labeling quality) and efficiency (inference speed) to comparatively study the performance of existing methods. We also present the performance results of various methods both in terms of accuracy and efficiency and compare them based on the key strategies employed in the frameworks.