PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLOv5, YOLOv8 and YOLOv10: The Go-To Detectors for Real-time Vision

Muhammad Hussain

arXiv:2407.02988v1cs.CV

TL;DR

The review examines how YOLOv5, YOLOv8, and YOLOv10 advance real-time object detection for edge deployment. It compares their architectures and performance, finding a progression toward improved accuracy and efficiency, while emphasizing application-specific trade-offs.

  • Problem

    The review addresses how successive YOLO variants balance speed, accuracy, efficiency, and hardware compatibility for resource-constrained edge deployment.

  • Method

    The paper reviews YOLOv5, YOLOv8, and YOLOv10, highlighting architectural advances, comparing performance metrics, and discussing edge-deployment suitability.

  • Results

    The reviewed variants show progressive improvements in real-time detection architecture, accuracy, efficiency, and deployment capability, including YOLOv10's NMS-free inference and optimized feature processing.

  • Takeaways & Limitations

    YOLOv5, YOLOv8, and YOLOv10 are promising edge-deployment options, but selection should reflect application requirements, target hardware, and resource constraints.

  • Takeaways & Limitations

    The appropriate YOLO variant depends on the specific application requirements, target hardware, and balance between performance and resource constraints.

Abstract

from arXiv · show

This paper presents a comprehensive review of the evolution of the YOLO (You Only Look Once) object detection algorithm, focusing on YOLOv5, YOLOv8, and YOLOv10. We analyze the architectural advancements, performance improvements, and suitability for edge deployment across these versions. YOLOv5 introduced significant innovations such as the CSPDarknet backbone and Mosaic Augmentation, balancing speed and accuracy. YOLOv8 built upon this foundation with enhanced feature extraction and anchor-free detection, improving versatility and performance. YOLOv10 represents a leap forward with NMS-free training, spatial-channel decoupled downsampling, and large-kernel convolutions, achieving state-of-the-art performance with reduced computational overhead. Our findings highlight the progressive enhancements in accuracy, efficiency, and real-time performance, particularly emphasizing their applicability in resource-constrained environments. This review provides insights into the trade-offs between model complexity and detection accuracy, offering guidance for selecting the most appropriate YOLO version for specific edge computing applications.

1 Introduction

This review compares YOLOv5, YOLOv8, and YOLOv10 as prominent real-time detectors for resource-constrained edge deployment. It focuses on their architectural advances, performance, and deployment trade-offs.

  • YOLOv5, YOLOv8, and YOLOv10 are highlighted for balancing speed, accuracy, and efficiency in constrained environments.
  • YOLOv5 established a flexible edge-deployment option through modular design and export support for ONNX, CoreML, and TFLite.
  • YOLOv8 added anchor-free detection and a unified framework, improving versatility and performance on small objects.
  • YOLOv10 introduced NMS-free training and holistic model design to reduce computational overhead while maintaining high accuracy.
  • The review compares key advancements and performance metrics to guide YOLO selection for specific edge-computing requirements.

2 YOLOv5

YOLOv5 combines an efficient CSP-based architecture, multi-scale feature processing, flexible deployment, and size variants for different edge constraints. Its COCO results increase with model size, alongside higher latency and computational demand.

  • YOLOv5 uses CSP connections, spatial pyramid pooling, a PAN neck, and anchor-based predictions for multi-scale object detection.
  • YOLOv5 provides nano-to-extra-large variants, allowing deployment choices spanning severely constrained devices to accuracy-sensitive applications.
  • YOLOv5 supports PyTorch-based training and exports to TensorFlow, ONNX, CoreML, TensorRT, and other deployment formats.
  • Mosaic augmentation combines four training images to improve small-object detection and robustness across scales and aspect ratios.
  • Summary of Metrics: YOLOv5x reaches 50.7% COCO AP (val) with 766 ms CPU latency, 86.7 million parameters, and 205.7 billion FLOPs.

3 YOLOv8

YOLOv8 extends YOLOv5 with stronger feature extraction and fusion, anchor-free detection, training efficiencies, and a unified multi-task framework. Its larger variants deliver higher COCO AP with increased latency and FLOPs.

  • PANet Neck: PANet combines bottom-up and top-down paths to improve information flow, multi-scale feature fusion, and small-object detection.
  • Enhanced Post-processing: The anchor-free detection head simplifies architecture, reduces computational overhead, and supports detection of small, densely packed objects.
  • Training Efficiency: YOLOv8 uses mixed-precision training, automated hyperparameter optimization, and a unified framework for detection, segmentation, and pose estimation.
  • Architecture Improvements: YOLOv8 introduces the C2f block and enhanced CSPDarknet design to improve feature extraction, feature reuse, and model-size efficiency.
  • Summary of Metrics: Increasing YOLOv8 model size raises performance metrics while also increasing computational demands and latency.
  • Summary of Metrics: YOLOv8x achieves 53.9% COCO AP (val) with 479.1 ms ONNX CPU latency, 3.53 ms A100 TensorRT latency, and 257.8 billion FLOPs.

4 YOLOv10

YOLOv10 targets efficiency-accuracy trade-offs through NMS-free training and architectural optimizations. Its variants report low latency alongside strong AP, with improvements over corresponding YOLOv8 variants.

  • NMS-Free Training and Inference: YOLOv10 uses consistent dual assignments to eliminate NMS during inference and reduce latency for real-time deployment.
  • Holistic Efficiency-Accuracy Driven Model Design: Spatial-channel decoupled downsampling, a lightweight classification head, and rank-guided blocks reduce computational overhead.
  • Enhanced Model Capabilities: YOLOv10 improves small-object detection and reduces false positives through enhanced feature extraction and classification.
  • Conclusion: YOLOv10 variants are presented as suitable for real-time applications requiring computational efficiency and accuracy.
  • Performance Comparison: YOLOv10-N achieves 39.5% AP with 1.84 ms latency, using 2.3 million parameters and 6.7 GFLOPs.
  • Performance Comparison: YOLOv10-X achieves 54.4% AP with 10.70 ms latency, compared with YOLOv8-X’s 53.9% AP and 16.86 ms latency.

5.1 Architectural Features Comparison

YOLOv5, YOLOv8, and YOLOv10 show progressively optimized backbones, PANet-based feature fusion, and architectural designs aimed at balancing computational cost with feature-extraction efficiency.

  • Comparison: Table 4 presents a comparison of key architectural features across YOLOv5, YOLOv8, and YOLOv10.The table is described as a detailed architectural-features comparison.
  • Backbone: YOLOv5 introduced CSPDarknet, while YOLOv8 and YOLOv10 further optimized its feature-extraction capabilities.The progression includes improved convolution operations, channel configurations, and potentially rank-guided block design.
  • Feature Fusion: YOLOv5 used PANet for multi-scale feature fusion, followed by proposed improvements in YOLOv8 and efficiency optimizations in YOLOv10.YOLOv10's PANet changes likely incorporate spatial-channel decoupled operations for feature propagation.

5.1.3 Detection Head Design

The detection head evolves from YOLOv5's anchor-based design to anchor-free designs in YOLOv8 and YOLOv10, with YOLOv10 adding dual assignments for NMS-free operation.

  • Detection Head Design: YOLOv5 used predefined anchor boxes for object detection.Its detection head followed an anchor-based approach.
  • Detection Head Design: YOLOv8 shifted to an anchor-free design, simplifying the detection process and potentially improving performance on small objects.The passage presents the small-object benefit as potential rather than established.

5.1.4 Non-Maximum Suppression (NMS)

YOLOv5 and YOLOv8 required NMS to remove redundant detections, whereas YOLOv10 introduced NMS-free training and inference to reduce deployment overhead and latency.

  • NMS Requirement: YOLOv5 and YOLOv8 required NMS as post-processing to filter redundant detections.NMS was applied after the model generated detections.
  • NMS-Free Design: YOLOv10 introduced NMS-free training and inference, significantly reducing computational overhead and deployment latency.The change removes NMS from both training and inference.
  • Activation Functions: YOLOv5 used Leaky ReLU, while YOLOv8 adopted SiLU activation for smooth gradients and potential performance benefits.These activation choices distinguish the two versions' implementation details.
  • Feature Pyramid Network: YOLOv10 enhanced FPN with spatial-channel decoupling for potentially more efficient and effective multi-scale feature processing.The passage characterizes the benefits as potential outcomes of the architectural modification.

5.1.7 Loss Function

The loss-function progression moves from YOLOv5's CIoU loss to task-specific losses in YOLOv8 and YOLOv10's Consistent Dual Assignment loss aligned with NMS-free training.

  • Loss Function: YOLOv5 employed CIoU loss to balance bounding-box regression and classification.CIoU is presented as the version's loss design.
  • Loss Function: YOLOv8 introduced task-specific losses to optimize performance for different computer-vision tasks.The loss design was adapted to task requirements.
  • Loss Function: YOLOv10 developed Consistent Dual Assignment loss in alignment with its NMS-free training approach.The passage also identifies potential overall detection-accuracy benefits.

5.1.8 Data Augmentation

The reviewed YOLO variants use different augmentation and training strategies, with YOLOv10 combining dual assignments and two-stage training. These design choices are associated in the review with refined feature learning and improved detection performance.

  • 5.1.8 Data Augmentation: YOLOv5 uses Mosaic and Cutout augmentations to enhance model robustness.
  • 5.1.8 Data Augmentation: YOLOv8 adds Mixup to its augmentation pipeline to further improve generalization.
  • 5.1.8 Data Augmentation: YOLOv5 and YOLOv8 employ single-stage training, consistent with typical YOLO models.
  • 5.1.8 Data Augmentation: YOLOv10 adopts two-stage training with dual assignments, potentially enabling more refined feature learning and improved detection performance.
  • 5.1.8 Data Augmentation: YOLOv10’s NMS-free training and spatial-channel decoupling contribute to improved accuracy, reduced computational overhead, and enhanced real-time performance.

6 Conclusion

The conclusion presents YOLOv5, YOLOv8, and YOLOv10 as progressively more accurate, efficient, and deployment-friendly detectors. YOLOv10 shows the strongest gains, especially in smaller models, while model selection remains dependent on application requirements and target hardware.

  • 6.1 Performance and Accuracy: YOLOv5 established an efficient foundation with CSPDarknet and PANet, balancing speed and accuracy.
  • 6.1 Performance and Accuracy: YOLOv8 enhanced feature extraction and introduced an anchor-free detection head, improving small-object performance and task versatility.
  • 6.2 Architectural Innovations: YOLOv10 uses NMS-free training and inference, spatial-channel decoupled downsampling, rank-guided blocks, large-kernel convolutions, and partial self-attention.
  • 6.4 Common Themes and Trends: The comparison identifies a consistent reduction in parameters and FLOPs while maintaining or improving accuracy, supporting edge deployment.
  • 6.4 Common Themes and Trends: All three variants offer multiple model sizes, enabling choices that match performance requirements and hardware constraints.
  • 6.4 Common Themes and Trends: All three variants support real-time inference, including their largest models when paired with appropriate hardware.
  • 6.5 Final Thoughts: YOLOv10 outperforms its predecessors particularly in smaller models, making it highly suitable for edge deployment where accuracy and efficiency are critical.
  • 6.5 Final Thoughts: Variant selection should consider application requirements, target hardware, and the balance between performance and resource constraints.
Loading 2407.02988v1…