Hierarchy is a fundamental umbrella concept in visual systems used in different ways in Computer Vision. It is based on relations of various types between the elements that make up a scene, including generalization/specialization and composition/decomposition. In this work, we assume that combining these different relations can improve object detection perfomance. We propose a lightweight integration of hierarchy during both training and inference within an object detection Deep Learning architecture, specifically in the YOLO (You Only Look Once) 14 toolset. We introduce YOLO-KM, an extension of YOLO that associates sets of classes derived from the knowledge models referred to as variants with objects rather than simple classes. This contribution also addresses the adaptation of annotations and data processing across all stages of training and inference. We conduct a comparative study between standard YOLO and YOLO-KM for object and part detection, based on the widely used Pascal VOC, Pascal Part, and Semantic Pascal Part datasets. The results demonstrate that our approach improves YOLO's detection performance without compromising training or inference efficiency. Our work is available in a GitHub repository : https://github.com/ChristophePantel/ultralytics/tree/hierarchical.
Pantel et al. (Thu,) studied this question.