Uncertainty Region Discovery and Model Refinement for Domain Adaptation in Road Detection

Abstract

Road detection is an important task in intelligent transportation systems. In regard to the domain adaptation, although self-training methods generate pseudo-labels for retraining the model, redundancy and noise in pseudo-labels lead to limited improvement. We propose that necessary annotations are required to effectively handle this challenge. First, we introduce the classifier discrepancy to discover and annotate uncertainty regions in the target domain. Then, we also design a recurrent teacher–student module to consider both prior knowledge and correction signals, avoiding the risk of suboptimal solution entrapment. Experiments on public data sets show that our approach is competitive with state-of-the-art approaches.