Active Reinforcement Learning for the Semantic Segmentation of Urban Images

Abstract

Image segmentation using supervised learning algorithms usually requires large amounts of annotated training data, while urban datasets frequently contain unbalanced classes leading to poor detection of under-represented classes. We investigate the use of a reinforced active learning method to address the limitations of semantic segmentation on complex urban scenes. In this method, an agent learns to select small informative regions of the image to be labeled from a pool of unlabeled data. The agent is represented by a deep Q-Network, where a Markov Decision Process (MDP) is used to formulate the Active Learning problem. We introduced the Frequency Weighted Average IoU (FWA IoU) as the image region selection performance metric to reduce the amount of training data while achieving competitive results. Using the Cityscapes and GTAv urban datasets, three baseline image segmentation networks (FPN, DeepLabV3, DeepLabV3+) trained with image regions selected by the proposed FWA IoU metric performed better compared to baseline region selection by active learning methods such as the Random selection, Entropy-based selection, and Bayesian Active Learning by Disagreement. Training performance equivalent to 98% of the fully supervised segmentation network was achieved by labeling only 8% of the dataset.