Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

hYOLO Model: Enhancing Object Classification with Hierarchical Context in YOLOv8

Veska Tsenkova, Peter Stanchev, Daniel Petrov, Deyan Lazarov

TL;DR

This work introduces hYOLO, a hierarchical end-to-end model built on YOLOv8 to exploit multi-level class taxonomies and visual similarity. It contributes a hierarchical architecture integrated into the YOLO detection head, a hierarchy-aware loss that penalizes non-child predictions, and a hierarchical evaluation metric that favors semantically closer mistakes. The authors evaluate two hierarchical structures—one taxonomy-based and one visually informed—on grocery-item datasets, demonstrating improved hierarchical accuracy, better calibration of predictions, and meaningful reductions in semantically distant misclassifications compared with flat YOLOv8. The approach offers practical benefits for real-world systems in retail, healthcare, and automated surveillance by enabling context-aware recognition and controlled error severity, with potential for deeper hierarchies and broader domain adoption in future work.

Abstract

Current convolution neural network (CNN) classification methods are predominantly focused on flat classification which aims solely to identify a specified object within an image. However, real-world objects often possess a natural hierarchical organization that can significantly help classification tasks. Capturing the presence of relations between objects enables better contextual understanding as well as control over the severity of mistakes. Considering these aspects, this paper proposes an end-to-end hierarchical model for image detection and classification built upon the YOLO model family. A novel hierarchical architecture, a modified loss function, and a performance metric tailored to the hierarchical nature of the model are introduced. The proposed model is trained and evaluated on two different hierarchical categorizations of the same dataset: a systematic categorization that disregards visual similarities between objects and a categorization accounting for common visual characteristics across classes. The results illustrate how the suggested methodology addresses the inherent hierarchical structure present in real-world objects, which conventional flat classification algorithms often overlook.

hYOLO Model: Enhancing Object Classification with Hierarchical Context in YOLOv8

TL;DR

This work introduces hYOLO, a hierarchical end-to-end model built on YOLOv8 to exploit multi-level class taxonomies and visual similarity. It contributes a hierarchical architecture integrated into the YOLO detection head, a hierarchy-aware loss that penalizes non-child predictions, and a hierarchical evaluation metric that favors semantically closer mistakes. The authors evaluate two hierarchical structures—one taxonomy-based and one visually informed—on grocery-item datasets, demonstrating improved hierarchical accuracy, better calibration of predictions, and meaningful reductions in semantically distant misclassifications compared with flat YOLOv8. The approach offers practical benefits for real-world systems in retail, healthcare, and automated surveillance by enabling context-aware recognition and controlled error severity, with potential for deeper hierarchies and broader domain adoption in future work.

Abstract

Current convolution neural network (CNN) classification methods are predominantly focused on flat classification which aims solely to identify a specified object within an image. However, real-world objects often possess a natural hierarchical organization that can significantly help classification tasks. Capturing the presence of relations between objects enables better contextual understanding as well as control over the severity of mistakes. Considering these aspects, this paper proposes an end-to-end hierarchical model for image detection and classification built upon the YOLO model family. A novel hierarchical architecture, a modified loss function, and a performance metric tailored to the hierarchical nature of the model are introduced. The proposed model is trained and evaluated on two different hierarchical categorizations of the same dataset: a systematic categorization that disregards visual similarities between objects and a categorization accounting for common visual characteristics across classes. The results illustrate how the suggested methodology addresses the inherent hierarchical structure present in real-world objects, which conventional flat classification algorithms often overlook.
Paper Structure (20 sections, 7 equations, 12 figures, 4 tables)

This paper contains 20 sections, 7 equations, 12 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Hierarchical Architectures
  5. Hierarchical Models Evaluation Metrics
  6. Modified Loss Function
  7. Experiments
  8. Experimental Variants of Hierarchical Architectures
  9. Datasets
  10. Dataset Hierarchical Structures
  11. Dataset Hierarchy Version 1:
  12. Dataset Hierarchy Version 2:
  13. YOLO Model and Hyperparameters
  14. Model Training and Evaluation
  15. Conclusion
  16. ...and 5 more sections

Figures (12)

  • Figure 1: Most popular types of classifiers.
  • Figure 2: Overview of YOLOv8 head and detection modules, adapted from https://github.com/RangeKing/DL-Diagram/tree/cv illustrating both the architecture and proposed insertion points for hierarchical classification layers.
  • Figure 3: Hierarchical architecture $V4$ implemented in YOLOv8 .
  • Figure 4: A tree structured hierarchy.
  • Figure 5: Hierarchical architectures $V1$, $V2$ and $V3$ implemented in YOLOv8 .
  • ...and 7 more figures