Deep Active Learning for Autonomous Perception

Abstract

Traditional supervised learning requires significant amounts of labeled training data to achieve satisfactory results. As autonomous perception systems collect continuous data, the labeling process becomes expensive and time-consuming. Active learning is a specialized semi-supervised learning strategy that allows a machine learning model to achieve high performance using less training data, thereby minimizing the cost of manual annotation. We explore active learning for autonomous vehicles, and propose a novel deep active learning framework for object detection and instance segmentation. We review prominent active learning approaches, study their performances in the aforementioned computer vision tasks, and perform several experiments using state-of-the-art R-CNN-based models for datasets in the self-driving domain. Our empirical experiments on a number of datasets reflect that active learning reduces the amount of training data required. We observe that early exploration with instance-rich training sets leads to good performance, and that false positives can have a negative impact if not dealt with appropriately. Furthermore, we perform a qualitative evaluation using autonomous driving data collected from Trondheim, illustrating that active learning can help in selecting more informative images to annotate.