Accelerating the Training of Convolutional Neural Networks for Image Segmentation with Deep Active Learning

摘要

Semantic segmentation is an important perception function for automated driving (AD), but training a deep neural network for the task using supervised learning requires expensive manual labelling. Active learning (AL) addresses this challenge by automatically querying and selecting a subset of the dataset to label with the aim to iteratively improve the model performance while minimizing labelling costs. This paper presents a systematic study of deep AL for semantic segmentation and offers three contributions. First, we compare six different state-of-the-art querying methods, including uncertainty-estimate, Bayesian, and out-of-distribution methods. Our comparison uses the state-of-the-art image segmentation architecture DeepLab on the Cityscapes dataset. Our results demonstrate subtle differences between the querying methods, which we analyze and explain. We show that the differences are nevertheless robust by reproducing them on architecture-independent randomly generated data. Second, we propose a novel way to aggregate the output of a query, by counting the number of pixels having acquisition values above a certain threshold. Our method outperforms the standard averaging approach. Finally, we demonstrate that our findings remain consistent for whole images and image crops.