[논문 리뷰] What is YOLOv9: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector

Muhammad Yaseen|arXiv (Cornell University)|2024. 09. 12.

Industrial Vision Systems and Defect Detection인용 수 34

한 줄 요약

본 논문은 YOLOv9의 아키텍처(GELAN 및 PGI), 학습 방법, 및 성능을 분석하여 YOLOv8 대비 개선점을 보여주고 다양한 배포를 위한 모델 변형을 자세히 설명한다.

ABSTRACT

This study provides a comprehensive analysis of the YOLOv9 object detection model, focusing on its architectural innovations, training methodologies, and performance improvements over its predecessors. Key advancements, such as the Generalized Efficient Layer Aggregation Network GELAN and Programmable Gradient Information PGI, significantly enhance feature extraction and gradient flow, leading to improved accuracy and efficiency. By incorporating Depthwise Convolutions and the lightweight C3Ghost architecture, YOLOv9 reduces computational complexity while maintaining high precision. Benchmark tests on Microsoft COCO demonstrate its superior mean Average Precision mAP and faster inference times, outperforming YOLOv8 across multiple metrics. The model versatility is highlighted by its seamless deployment across various hardware platforms, from edge devices to high performance GPUs, with built in support for PyTorch and TensorRT integration. This paper provides the first in depth exploration of YOLOv9s internal features and their real world applicability, establishing it as a state of the art solution for real time object detection across industries, from IoT devices to large scale industrial applications.

연구 동기 및 목표

• Evaluate YOLOv9’s architectural innovations (GELAN and PGI) and their impact on gradient flow and feature extraction.
• Assess training methodologies (augmentation, loss, mixed precision) and their role in performance and efficiency.
• Compare YOLOv9 variants against YOLOv8 and benchmarks on MS COCO to guide deployment choices.
• Demonstrate practical deployment considerations including PyTorch and TensorRT integration.
• Provide guidance on annotation format and labeling tools compatible with YOLOv9.

제안 방법

• Introduce Programmable Gradient Information (PGI) to address gradient flow and information bottlenecks.
• Incorporate Generalized Efficient Layer Aggregation Network (GELAN) to enhance multi-scale feature aggregation.
• Maintain anchor-free bounding box prediction with reversible data paths enabled by PGI.
• Utilize mosaic and mixup data augmentation with mixed-precision training.
• Offer model variants (t, s, m, c, e) with corresponding parameter counts and accuracy figures.
• Provide evaluation on MS COCO and compare against YOLOv8 across metrics.

Figure 1: PGI Architecture in YOLOv9 [ 15 ]

실험 결과

연구 질문

• RQ1 How do GELAN and PGI affect gradient flow and feature fusion in YOLOv9 compared to prior YOLO versions?
• RQ2 What are the trade-offs between model size, speed, and accuracy across YOLOv9 variants on MS COCO?
• RQ3 How does YOLOv9 performance (mAP, inference time) compare to YOLOv8 and other baselines?
• RQ4 What deployment workflows (PyTorch, TensorRT) are enabled by YOLOv9 for edge to server environments?
• RQ5 What annotation formats and labeling tools best integrate with YOLOv9 workflows?

주요 결과

• YOLOv9 achieves a 49% reduction in parameters and a 43% reduction in computation versus YOLOv8 with a 0.6% mAP improvement on MS COCO.
• YOLOv9 variants span from lightweight edge models to high-accuracy counterparts (t, s, m, c, e) with corresponding parameter counts and mAP.
• Table comparisons show mAP@0.5 of 53% (YOLOv7 AF) to 72.8% (YOLOv9-E) and inference times down to 23 ms on the tested setup.
• YOLOv9t and YOLOv9s target resource-constrained environments; YOLOv9e achieves the highest accuracy (55.6% mAP) with substantial parameter efficiency.
• GELAN and PGI address information bottlenecks and vanishing gradients, enabling lightweight models to achieve strong accuracy.
• YOLOv9 supports PyTorch and TensorRT, facilitating real-time deployment across edge to GPU platforms.

Figure 2: GELAN Architecture in YOLOv9 [ 16 ]