Hyper-YOLO: When Visual Object Detection Meets Hypergraph Computation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional object detection models like YOLO have limitations in capturing high-order feature interrelationships due to neck design constraints.
• Integrating cross-level features and complex correlations is crucial for advancing object detection accuracy.

Purpose of the Study:

• To introduce Hyper-YOLO, a novel object detection method that leverages hypergraph computations for enhanced feature correlation analysis.
• To address the limitations of existing YOLO models in exploiting high-order feature interrelationships and cross-level feature integration.

Main Methods:

• Proposed the Hypergraph Computation Empowered Semantic Collecting and Scattering (HGC-SCS) framework to transpose visual features into a semantic space and construct hypergraphs.
• Incorporated the Mixed Aggregation Network (MANet) in the backbone for improved feature extraction.
• Introduced the Hypergraph-Based Cross-Level and Cross-Position Representation Network (HyperC2Net) in the neck for sophisticated high-order interactions across multiple scales and positions.

Main Results:

• Hyper-YOLO demonstrates superior performance on the COCO dataset, positioning it as a state-of-the-art architecture.
• Hyper-YOLO-N achieved significant improvements, outperforming YOLOv8-N by 12% and YOLOv9-T by 9%.

Conclusions:

• Hyper-YOLO effectively captures complex high-order correlations among visual features by integrating hypergraph computations.
• The proposed HGC-SCS framework, MANet, and HyperC2Net enable advanced semantic and structural information acquisition, surpassing conventional feature-focused learning.