Enhancing Vehicle Type Classification in Rare Scenarios with Consistency-Aware Generation

Vehicle type classification (VTC) has broad application prospects in many aspects such as smart cities and unmanned driving. Using a Deep Neural Network (DNN) for VTC has made significant progress, but current methods have primarily focused on common scenarios, with insufficient attention given to rare scenarios, such as scenarios with heavy rain. Due to the difficulty in collecting vehicle images in rare scenarios, existing methods have difficulty achieving satisfactory results. In this paper, we propose a novel method for enhancing VTC in rare scenarios. To improve model performance in these challenging conditions, we firstly employ a subject-driven generative model to create images of specified vehicles in rare scenarios, addressing the issue of data scarcity. However, the generative model may produce noisy samples with inconsistent vehicles. We further introduce a salient detection model to eliminate the interference of the background in the samples, and subsequently reuse the classification model from common scenarios to score the samples for consistency, thereby guiding the training process. Extensive experiments have shown that our proposed method can effectively improve the performance of various classifiers in rare scenarios.