Road Scene Data Annotation with Semi-Automated Active Learning Framework for Convolutional Neural Networks

Autonomous driving vehicles are considered the future of mobility as they can reduce the mortality rate owing to traffic accidents. This can also be achieved using cameras and a Convolutional Neural Network (CNN) to detect objects on the road and take necessary actions to prevent life-threatening occurrences. However, the current form of CNN needs to be trained using large amounts of annotated data, which is time consuming, expensive, and requires extensive manpower. These limitations can be overcome by using Active Learning (AL) systems, which only select a subset of informative data from the big data for annotation by humans. Although AL reduces the amount of data being used for CNN training, humans are still needed to annotate the data. This study proposes a Semi-Automated Active Learning system (SAAL) to further reduce the need for manpower for data annotation. SAAL uses AL and a new algorithm called Machine Teachers (MTs), which are stacked algorithms of pre-trained CNN and optical flow that use the temporal-spatial information video data from cameras on vehicles to help humans annotate images. This allows SAAL to be partially automated and further reduces human effort while roughly maintaining the accuracy of CNN to that of AL.