A YOLO Analysis for Vehicle Recognition - Detection Improvements, Cropping Strategies and Cascading Architecture

Abstract

With the development of high-resolution cameras and enhanced storage devices, image datasets are increasingly captured with higher spatial dimensions. Many modern object detector architectures, such as the YOLO family, typically operate at a 640x640 resolution. When training on high-resolution images, the rescaling process can drastically reduce the region of interest, especially for low- or medium-sized objects far from the camera. To analyze the impact of rescaling, a six-vehicle dataset was constructed from high-resolution images. An automated preprocessing pipeline is built, that crops each car region, enforcing a minimum size of 640x640 by symmetrically expanding the margins. A refactoring process was applied to the new set of images, resulting in a new dataset centered on the region of interest. A YOLOv10 model was trained for each dataset, with the cropped dataset achieving a higher mAP50 of 0.995 compared to the uncropped version's 0.987. A cascading multilevel model employing both YOLO models was further proposed, with the first model analyzing the initial high-resolution image at a coarser level, cropping the vehicle area, and sending it to the second YOLO model for fine-level analysis. This architecture highlights the importance of preserving fine, detailed features that would otherwise be lost during scaling.