[논문 리뷰] What is YOLOv5: A deep look into the internal features of the popular object detector

Rahima Khanam, Muhammad Azhar Hussain|arXiv (Cornell University)|2024. 07. 30.

COVID-19 diagnosis using AI인용 수 27

한 줄 요약

이 논문은 YOLOv5 아키텍처, 학습 방법, 및 성능을 분석하고 CSP 백본, PA-Net 넥, 데이터 증강, PyTorch 전환을 상세히 다루며 모델 패밀리와 엣지 배치에 대한 함의를 제시한다.

ABSTRACT

This study presents a comprehensive analysis of the YOLOv5 object detection model, examining its architecture, training methodologies, and performance. Key components, including the Cross Stage Partial backbone and Path Aggregation-Network, are explored in detail. The paper reviews the model's performance across various metrics and hardware platforms. Additionally, the study discusses the transition from Darknet to PyTorch and its impact on model development. Overall, this research provides insights into YOLOv5's capabilities and its position within the broader landscape of object detection and why it is a popular choice for constrained edge deployment scenarios.

연구 동기 및 목표

• Assess YOLOv5 performance versus state-of-the-art object detectors across variants (n, s, m, l, x).
• Identify architectural innovations (CSP backbone, PA-Net neck) and training techniques contributing to efficiency and accuracy.
• Evaluate the impact of data augmentation, loss design, and 16-bit precision on real-time detection.
• Discuss implications of transitioning from Darknet to PyTorch for development and deployment.

제안 방법

• Describe the evolution and architectural footprint of YOLOv5, including backbone, neck, and head components.
• Detail training methodologies: data augmentation (including mosaic), loss components (GIoU/CIoU, BCE for classification and objectness).
• Explain transition from Darknet to PyTorch and its impact on development and deployment.
• Explain data augmentation and anchor box strategy for bounding box prediction.
• Present 16-bit floating point precision implications for inference speed on specific GPUs.
• Outline CSP backbone and PA-Net neck designs and their roles in efficiency.

Figure 1: Process of Object Detection [ 13 ]

실험 결과

연구 질문

• RQ1 How do YOLOv5 variants (n, s, m, l, x) compare in accuracy and speed across CPU and GPU platforms?
• RQ2 What architectural choices (CSP backbone, PA-Net neck) and training techniques drive YOLOv5 performance gains?
• RQ3 What is the impact of transferring YOLOv5 from Darknet to PyTorch on development and deployment?
• RQ4 How do data augmentation and bounding box anchor strategies affect detection of small objects and overall mAP?

주요 결과

• YOLOv5 variants show increasing mAP with more parameters, balanced by inference speed across CPU/GPU.
• CSP backbone and PA-Net neck contribute to improved efficiency without sacrificing accuracy.
• Mosaic augmentation enhances small-object detection and generalization.
• 16-bit precision can speed up inference on certain GPUs (V100, T4) without universal hardware support.
• Transition to PyTorch democratizes access and deployment, expanding practical adoption.

Figure 2: Bounding box prediction based on an anchor box [ 15 ]