CHARD: Characteristic-Aware Hierarchical Detection for Bangladesh Road Scenes

Abstract

Traffic in South Asia is dense, weakly structured, and includes vehicle types such as auto-rickshaws, pedal-powered three-wheelers, and cart vehicles that are absent from standard benchmarks such as COCO, BDD100K, and Waymo. Detection is further challenged by severe class imbalance, where the locally important vehicle classes are also the rarest. We propose Characteristic-Aware Hierarchical Detection (CHARD), which uses vehicle characteristics, wheel count, size, and propulsion as auxiliary supervision during training. This allows rare classes to learn from shared attributes instead of relying only on limited class-specific examples. CHARD is implemented as an end-to-end hierarchical attribute-to-class head for RT-DETR and as a lightweight auxiliary attribute branch for YOLO that adds no parameters or inference latency. We evaluate 72 independently trained models (eight variants, three random seeds) on three Bangladeshi road-vehicle datasets: BadODD, Poribohon-BD, and Sorokh-Poth. On BadODD, CHARD with a YOLOv11-l backbone achieves the highest seed-averaged test mAP@50:95 of 44.3%, outperforming its no-attribute counterpart by 2.4 points ( $p=0.003$ ) and the standard YOLOv11-l baseline by 1.8 points. The improvement is concentrated on tail classes, where AP increases from 28.0% to 31.9%, while head and medium-frequency classes remain unchanged. CHARD also matches or exceeds the baseline across all frequency groups on both validation and test sets without increasing inference cost. The benefit is conditional rather than universal. Auxiliary supervision is statistically neutral on the lower-capacity YOLOv8-l backbone and on the two supplementary datasets with weaker class imbalance. These results suggest that characteristic-aware supervision is most effective when sufficient model capacity and pronounced long-tail distributions allow shared attributes to improve rare-class learning.