Weakly-supervised road condition detection via scribble annotations

Abstract

Deep learning-based solutions in road condition detection rely on fully supervised learning widely which requires a large quantity of training data with pixel-wise annotation from researchers. Accurate dense labeling is challenging since the boundaries between road condition regions are ambiguous. To combat this issue, this work aims to introduce a weakly supervised road condition detection framework, which can generate high-quality pseudo masks from sparse scribble labels and train a road condition detection network using these masks. Specifically, we collect a dataset for road condition detection and annotate it with scribble. Next, we propose a graph convolutional network (GCN)-based label augmentation strategy, which considers both local and global image information, to generate pixel-level pseudo-labels by augmenting the label information from scribbles to the whole. To alleviate the label inconsistency caused by sparse annotations, we adopt the supervision strategy with joint loss of labeled and unlabeled regions during training. Extensive experiments demonstrate that the proposed method can work well on various road condition detection and is on par with the full-supervision method. The code will be made publicly available at https://github.com/qinhs9/Scribble_road_Condition.