Authors: Parag Hossain
Abstract: Urban vehicle detection systems face a fundamental challenge that is often overlooked in benchmark datasets: severe class imbalance. In real-world traffic scenes, common vehicles such as cars appear thousands of times more frequently than critical but rare categories including ambulances, e-bikes, and motorcycles. This imbalance causes standard detectors to become biased toward majority classes, leading to unacceptable failure rates for minority class detection in safety-critical applications. In this paper, we propose a novel Class-Balanced Knowledge Distillation (CBKD) framework specifically designed to address this challenge on the challenging CAVI-14 dataset, which contains fourteen urban vehicle categories with up to fifteen-fold class imbalance. Our method integrates three key components: class-balanced sampling to ensure equal exposure to all classes during training, focal loss with class-specific weights to down-weight easy majority examples, and knowledge distillation from a teacher model pretrained on a synthetically balanced dataset. Extensive experiments demonstrate that CBKD achieves perfect mean average precision at 0.50 intersection-over-union threshold (mAP50) of 1.000 and near-perfect mAP50-95 of 1.000 after one thousand training epochs. Per-class F1 scores consistently exceed 0.97 across all fourteen categories, including the rarest classes. Qualitative results on validation images show accurate detection even under heavy occlusion and challenging lighting conditions. Our approach establishes a new state-of-the-art on the CAVI-14 dataset and provides a practical, reproducible solution for imbalanced object detection in intelligent transportation systems.
DOI: https://doi.org/10.5281/zenodo.20351972
