Active Reinforcement Learning for the Semantic Segmentation of Images Captured by Mobile Sensors

Neural Networks have been employed to attain acceptable performance on semantic segmentation. To perform well, many supervised learning algorithms require a large amount of annotated data. Furthermore, real-world datasets are frequently severely unbalanced, resulting in poor detection of underrepresented classes. The annotation task requires time-consuming human labor. This thesis investigates the use of a reinforced active learning as region selection method to reduce human labor while achieving competitive results. A Deep Query Network (DQN) is utilized to identify the best strategy to label the most informative regions of the image. A Mean Intersection over Union (MIoU) training performance equivalent to 98% of the fully supervised segmentation network was achieved with labeling only 8% of dataset. Another 8% of labelled dataset was used for training the DQN. The performance of all three segmentation networks trained with regions selected by Frequency Weighted Average (FWA) IoU is better in comparison with baseline methods.