PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLO advances to its genesis: a decadal and comprehensive review of the You Only Look Once (YOLO) series

Ranjan Sapkota, Marco Flores Calero, Rizwan Qureshi, Chetan Badgujar, Upesh Nepal, Alwin Poulose, Peter Zeno, Uday Bhanu Prakash Vaddevolu, Sheheryar Khan, Maged Shoman, Hong Yan, Manoj Karkee

arXiv:2406.19407v8cs.CV

TL;DR

The review addresses the need for an updated, systematic account of YOLO’s newer iterations and application domains. It analyzes YOLO in reverse chronological order, compares versions and alternatives, and reports a decade of developments in speed, accuracy, efficiency, applications, and limitations. The review concludes by outlining a prospective move toward multimodal and context-aware YOLO systems.

  • Problem

    Existing reviews may omit YOLOv9 through YOLOv12 or application domains, motivating an updated comparative account of the YOLO family.

  • Method

    The review analyzes YOLO from YOLOv12 backward to YOLOv1, compares versions and alternatives, and examines applications, performance enhancements, and challenges.

  • Results

    The review chronicles a decade of YOLO advancements in detection speed, accuracy, computational efficiency, applications, and version-specific limitations.

  • Takeaways & Limitations

    The review projects YOLO’s integration with multimodal data, large language models, and natural language processing toward more context-aware applications.

  • Takeaways & Limitations

    YOLOv12’s attention mechanisms reduce inference speed to 30 FPS versus YOLO11’s 40 FPS and increase training time by approximately 20%.

Abstract

from arXiv · show

This review systematically examines the progression of the You Only Look Once (YOLO) object detection algorithms from YOLOv1 to the recently unveiled YOLOv12. Employing a reverse chronological analysis, this study examines the advancements introduced by YOLO algorithms, beginning with YOLOv12 and progressing through YOLO11 (or YOLOv11), YOLOv10, YOLOv9, YOLOv8, and subsequent versions to explore each version's contributions to enhancing speed, detection accuracy, and computational efficiency in real-time object detection. Additionally, this study reviews the alternative versions derived from YOLO architectural advancements of YOLO-NAS, YOLO-X, YOLO-R, DAMO-YOLO, and Gold-YOLO. Moreover, the study highlights the transformative impact of YOLO models across five critical application areas: autonomous vehicles and traffic safety, healthcare and medical imaging, industrial manufacturing, surveillance and security, and agriculture. By detailing the incremental technological advancements in subsequent YOLO versions, this review chronicles the evolution of YOLO, and discusses the challenges and limitations in each of the earlier versions. The evolution signifies a path towards integrating YOLO with multimodal, context-aware, and Artificial General Intelligence (AGI) systems for the next YOLO decade, promising significant implications for future developments in AI-driven applications. YOLO Review, YOLO Advances, YOLOv13, YOLOv14, YOLOv15, YOLOv16, YOLOv17, YOLOv18, YOLOv19, YOLOv20, YOLO review, YOLO Object Detection

1 Introduction

Object detection evolved from hand-crafted features and classical classifiers toward deep learning methods that better address the variability of real-world scenes. This progression led to proposal-based detectors and unified approaches such as YOLO, which combine localization and classification for real-time use.

  • Real-world motivation: Real-time object detection supports monitoring and interaction in dynamic environments across agriculture, transportation, education, and healthcare.It is foundational for autonomous vehicles and robotic systems, where rapid recognition and tracking of vehicles, pedestrians, bicycles, and obstacles supports navigational safety and efficiency.
  • Real-world motivation: Variable orientations, scales, distances, lighting, shadows, and glare make robust real-time detection difficult in real-world environments.These conditions challenge an algorithm’s ability to generalize while maintaining accuracy.
  • Traditional approaches: Traditional detectors used hand-crafted features, including Gabor features, HOG, LBP, Haar-like features, DPM, and SIFT, with classifiers such as SVM, Bayesian classifiers, AdaBoost, Random Forest, and MLPs.These approaches offered moderate success under controlled conditions but struggled with robustness and generalization in diverse scenes.
  • Deep learning detectors: CNNs learn hierarchical and spatially invariant representations, but classification-oriented CNNs cannot directly handle varying numbers, sizes, aspect ratios, and orientations of detected objects.Object detection additionally requires localizing each object with bounding boxes.
  • Deep learning detectors: R-CNN generated about 2,000 region proposals per image, extracted CNN features for each, and classified regions with linear SVMs.Redundant computation across overlapping proposals motivated later improvements in detection efficiency.
  • Deep learning detectors: Fast R-CNN integrated proposal feature extraction and classification into one pass, while Faster R-CNN introduced region proposal networks for end-to-end training.These changes reduced redundant computation and dependence on external proposal generation.

2 The evolution of YOLO: trajectory and variants

YOLO evolved from a single-stage real-time detector into a family of increasingly efficient and capable models, culminating in attention-centric YOLOv12. Across versions, development emphasized accuracy, speed, scalability, and computational efficiency.

  • Foundations: YOLOv1 introduced unified single-stage detection, dividing images into grids and predicting bounding boxes and class probabilities directly for end-to-end learning.
  • Foundations: YOLOv2 expanded operating resolution and detection capacity to more than 9000 object categories, with smaller and larger variants balancing speed and accuracy.
  • Intermediate evolution: YOLOv3 added multi-scale predictions and a deeper network to improve small-object detection, while offering variants with different speed–accuracy trade-offs.
  • Intermediate evolution: YOLOv6 through YOLO11 continued improving scalability, computational demands, and real-time performance through architectural and training refinements.YOLOv6 introduced Anchor-Aided Training and Self-Distillation, while YOLOv7 used trainable bag-of-freebies and dynamic label assignment.
  • Recent versions: YOLOv9 introduced programmable gradient information and GELAN, whereas YOLOv10 removed NMS and optimized model components for efficient real-time detection.
  • Recent versions: 40.6% mAP at 1.64 ms was reported for YOLOv12-N on a T4 GPU, exceeding YOLOv10-N and YOLOv11-N by 2.1 mAP at comparable speed.YOLOv12 uses Area Attention and R-ELAN, supports detection, segmentation, classification, pose estimation, and oriented object detection, and offers five variants.

3 Prior YOLO literature: context and distinctions

Prior YOLO literature reviewed the algorithm’s evolution and applications but often omitted the newest iterations or selected application domains. This review addresses that gap through reverse-chronological technical and application-focused analysis.

  • Review context: The review collected published YOLO literature to document and critically analyze major developments, highlights, and limitations.
  • Existing literature: Earlier reviews examined YOLO evolution, comparisons with traditional CNNs, and practical impacts across application domains.
  • Existing literature: A manufacturing-focused review traced YOLOv1 through YOLOv8 and emphasized industrial defect detection.
  • Research gap: The proposed review targets omissions of YOLOv9 through YOLOv12 and extends coverage across multiple application domains.It aims to provide updated insights and comparative analysis for the research and technical community.
  • Review design: The analysis proceeds in reverse chronological order, grouping recent, intermediate, and earlier YOLO versions to trace incremental improvements.
  • Application scope: Applications are examined in five domains: autonomous vehicles and traffic safety, healthcare and medical image analysis, surveillance and security, industrial manufacturing, and agriculture.

4 Review of YOLO versions

The review presents YOLO models from YOLOv12 backward to YOLOv1. Starting with recent capabilities provides immediate insight into the state of the art before tracing the foundations of later innovations.

  • 4 Review of YOLO versions: The section begins with YOLOv12 and progressively traces the series back to YOLOv1.
  • 4 Review of YOLO versions: This ordering foregrounds recent technological advancements before explaining how earlier models laid the groundwork for subsequent innovations.

4.1 YOLOv12 and YOLO11

YOLOv12 advances real-time detection through attention-centered architecture, residual feature aggregation, and efficiency-oriented training and design choices. Reported results show improved accuracy, latency, and computational efficiency across model scales, while YOLO11 contributes streamlined feature extraction and reduced parameterization.

  • YOLOv12 and YOLO11: YOLOv12 reports lower latency and enhanced computational efficiency than prior YOLO versions, with YOLOv12-S running 42% faster than RT-DETR-R18 and RT-DETRv2-R18.The comparison states that YOLOv12-S uses 36% of the computation and 45% of the parameters of those detectors.
  • YOLOv12 and YOLO11: On COCO, YOLOv12-N achieves 40.6% mAP and YOLOv12-X reaches 55.2% mAP across the five-scale model family.YOLOv12-N reaches 40.6% mAP at 1.64 ms on a T4 GPU, while YOLOv12-S reaches 48.0% mAP at 2.61 ms/image.
  • YOLOv12 architectural innovation: YOLOv12 integrates Area Attention, R-ELAN blocks, and a streamlined detection head to optimize visual information processing while maintaining high accuracy.Area Attention uses segmented feature processing with Flash Attention, while R-ELAN combines residual shortcuts with dual-branch processing.
  • YOLOv12 architectural innovation: 50% lower computational complexity is reported for Area Attention through spatial reshaping while preserving large receptive fields.The module enables fixed-resolution real-time detection through optimized memory access patterns.
  • YOLOv12 architectural innovation: 18% fewer parameters and 24% fewer FLOPs are reported for R-ELAN’s final aggregation stage compared with baseline architectures.Residual connections and revised aggregation mitigate gradient blocking and support stable convergence, particularly in larger YOLOv12 models.
  • YOLOv12 and YOLO11: YOLO11 replaces C2f with C3k2, retains SPPF, and adds C2PSA attention to improve computational efficiency and feature extraction.These changes are described as improving precision for small or occluded objects; empirical evaluation reports 22% fewer parameters than YOLOv8m and inference approximately 2% faster than YOLOv10.

4.2 YOLOv10, YOLOv9 and YOLOv8

YOLOv10, YOLOv9, and YOLOv8 advance real-time detection through architectural changes targeting latency, information preservation, multi-scale representation, and task versatility.

  • YOLOv10: YOLOv10 removes non-maximum suppression and uses dual one-to-many and one-to-one label assignment to reduce latency while maintaining detection accuracy.Its architecture also includes lightweight classification heads, spatial-channel decoupled downsampling, and rank-guided block design.
  • YOLOv10: 54.4% mAP at 10.70 ms is reported for YOLOv10-X, while YOLOv10-N reaches 38.5% mAP at 1.84 ms.Intermediate configurations trade latency against mAP across Nano, Small, Medium, Balanced, Large, and Extra Large variants.
  • Evolution: Across the series, YOLOv9-X and YOLOv8-X report mAP values of 54.4% and 53.9%, while YOLOv10 is described as more efficient.The review frames these changes as incremental improvements in speed and accuracy for practical applications.
  • YOLOv9: YOLOv9 introduces Programmable Gradient Information and GELAN to mitigate information loss and improve efficiency and accuracy in deep processing.GELAN combines features across multiple depths and pathways, while related architectural descriptions emphasize flexible feature extraction.
  • YOLOv8: YOLOv8 combines a modified CSPDarknet53 backbone, C2f, and SPPF with multiple scaled configurations for varied computational and performance requirements.The SPPF layer accelerates computation by pooling features into a fixed-size map.

4.3 YOLOv7, YOLOv6 and YOLOv5

YOLOv7, YOLOv6, and YOLOv5 extend the series toward scale-robust detection, hardware-aware deployment, and production-oriented usability.

  • YOLOv7: YOLOv7 improves drone-captured detection through TPH and CBAM, emphasizing scale variation and densely packed objects in cluttered environments.Multi-scale testing and a self-trained classifier further refine the model’s detection process.
  • YOLOv6: YOLOv6 balances speed and accuracy for industrial deployment across hardware platforms through network design, training, and quantization techniques.Its scalable architecture ranges from YOLOv6-N to YOLOv6-X for different computational budgets.
  • YOLOv6: YOLOv6 introduces advanced label assignment techniques and loss functions to improve predictive accuracy and operational efficiency.The review also associates YOLOv6 with throughput and latency gains for real-time industrial and commercial applications.
  • YOLOv6 and YOLOv7: YOLOv6 adds lightweight nano models for mobile and CPU environments, while YOLOv7 introduces PANet and CIOU loss building on YOLOv5.These changes target small-object detection and feature aggregation across deployment settings.
  • YOLOv5: YOLOv5 emphasizes user-friendly, production-ready deployment through streamlined architecture, optimized feature extraction, and faster data throughput.Its design targets edge devices with limited computational resources while preserving detection accuracy.
  • YOLOv5: YOLOv5’s balance of speed, accuracy, and utility made it a preferred choice for efficient state-of-the-art detection deployment.

4.4 YOLOv4, YOLOv3, YOLOv2 and YOLOv1

YOLOv1 through YOLOv4 establish the series’ progression from unified single-shot detection toward multi-scale accuracy and computationally efficient architectures.

  • YOLOv4: YOLOv4 uses CSPDarknet-53 with Cross-Stage Partial connections to reduce computational demands while enhancing learning capacity.
  • YOLOv4: YOLOv4 combines Mish activation, Mosaic and CutMix augmentation, DropBlock, label smoothing, and BoF and BoS optimization strategies.These techniques target gradient behavior, regularization, training, and inference efficiency.
  • YOLOv3: YOLOv3 improves small-object detection by using three detection scales that capture features at different resolutions.
  • YOLOv2: YOLOv2 introduces a 30-layer architecture with Faster R-CNN anchor boxes and batch normalization to speed convergence and improve performance.
  • YOLOv1: YOLOv1 revolutionizes detection by predicting bounding boxes and class probabilities in one network pass through a unified grid-based architecture.Its simplified Darknet-19 backbone supports end-to-end object detection.

4.5 Alternative versions derived from YOLO

Alternative YOLO-derived models modify architecture, representation, feature fusion, quantization, and training to improve real-time detection across efficiency and accuracy goals.

  • Five primary alternatives are identified: YOLO-NAS, YOLOX, YOLOR, DAMO-YOLO, and Gold-YOLO.
  • YOLO-NAS: YOLO-NAS uses Neural Architecture Search and quantization-aware blocks to improve performance while minimizing precision loss after quantization.
  • YOLOX: YOLOX combines anchor-free detection, a decoupled head, and SimOTA label assignment to improve detection efficiency and convergence.
  • YOLOX: YOLOX-L reaches 50.0% AP on COCO at 68.9 FPS on Tesla V100, surpassing YOLOv5-L by 1.8% AP.
  • YOLOR: YOLOR unifies implicit and explicit knowledge representations for multi-task learning, achieving comparable accuracy to Scaled YOLOv4 while increasing inference speed by 88%.
  • DAMO-YOLO and Gold-YOLO: DAMO-YOLO integrates NAS, RepGFPN, lightweight heads, AlignedOTA, and distillation, while Gold-YOLO uses Gather-and-Distribute fusion to improve multi-scale information exchange.

5 Applications

YOLO models support real-time applications across transportation, healthcare, security, manufacturing, and agriculture, with the supplied examples emphasizing autonomous-vehicle and traffic-safety use cases.

  • The review organizes YOLO applications into autonomous vehicles and traffic safety, healthcare and medical imaging, security and surveillance, industrial manufacturing, and agriculture.
  • Autonomous Vehicles and Traffic Safety: In autonomous vehicles, YOLO supports obstacle detection, traffic-sign recognition, pedestrian analysis, and real-time environmental perception for navigation and safety.
  • Autonomous Vehicles and Traffic Safety: YOLO5-based systems are used for steering-angle prediction, occluded-object detection, and enhanced autonomous-driving detection, while YOLOv8 and YOLOv9 address adverse weather and small objects.
  • Pedestrian Analysis: YOLO combined with tracking and sequence models supports pedestrian crash-risk assessment, conflict prediction, crossing-intention analysis, and walking-speed measurement.
  • Traffic-Safety Systems: Traffic-safety systems also use YOLOv4 with OpenPose for crossing-intention and mobility-status detection, and YOLO-based systems for traffic-sign recognition and passenger counting.

6 Challenges and limitations

The review identifies persistent detection and deployment limitations across YOLO versions, especially for small, overlapping, rotated, or variably scaled objects. It also highlights continuing trade-offs among accuracy, speed, computational demand, generalization, and real-world validation.

  • YOLOv12: YOLOv12 improves architectural efficiency but may run at 30 FPS versus YOLO11’s 40 FPS, require about 20% longer training, and increase false positives at lower confidence thresholds.Its attention mechanisms have hardware-dependent benefits, complicating real-time deployment in resource-constrained environments.
  • YOLOv12 and earlier versions: YOLOv12 continues to struggle with small objects and crowded scenes because downsampling reduces fine-grained spatial information.These limitations persist from earlier YOLO iterations and affect dense detection scenarios.
  • YOLOv11: YOLOv11 remains vulnerable to small, low-resolution, and rotated objects, overfitting on limited datasets, and accuracy-efficiency trade-offs in complex environments.The review calls for stronger generalization and training approaches.
  • YOLOv10 and YOLOv9: YOLOv10 and YOLOv9 lack extensive validation and adoption across diverse applications, limiting confidence in their practical performance and comparisons with earlier versions.The review specifically notes limited published research for both versions.
  • YOLOv8 and earlier versions: YOLOv8 and earlier versions face deployment and detection constraints, including resource consumption, scale variability, motion or occlusion sensitivity, small-object errors, dataset inconsistency, and limited low-power suitability.YOLOv4 also shows inconsistencies with class imbalance and rare objects, while YOLOv1’s older limitations remain relevant to small-object localization.
  • Long-term version relevance: The review indicates that versions below YOLOv5 may gradually be displaced as newer models improve overall performance and efficiency.This reflects the series’ continuing replacement cycle rather than a complete removal from future research.

7 Future directions in object detection with YOLO

Future YOLO research targets deployment in increasingly constrained edge environments while extending interaction with physical systems. Efficiency improvements and embodied applications are highlighted as key directions.

  • Edge deployment: Future optimization should improve YOLO efficiency and accuracy on ultra-low-power microcontrollers and embedded systems.Suggested techniques include model pruning, quantization, and specialized hardware accelerators.
  • Edge deployment: YOLO deployment on edge devices creates opportunities for further research and development.
  • Embodied artificial intelligence: In embodied artificial intelligence, YOLO enhances sensory capabilities for interaction with the physical environment.Applications include autonomous vehicles, drones, robots, human-robot interaction, and robotic surgical assistants.

8 Expanding YOLO object detection into broader AI domains

The review frames YOLO as a visual component that can be integrated with broader artificial intelligence systems. Combining object detection with language understanding, reasoning, and contextual awareness is presented as a route toward more autonomous, real-time systems.

  • Artificial general intelligence: AGI systems would need YOLO-like object detection alongside natural language understanding, reasoning, and decision-making.This combination is described as supporting diverse tasks in dynamic environments and complex scenarios.
  • Artificial general intelligence: AGI is defined as an intelligent agent capable of solving complex problems across diverse domains at human-level or higher intelligence.
  • Large language model integration: Integrating YOLO with large language models could combine real-time visual interpretation with language understanding, reasoning, and contextual awareness.The proposed fusion would support object analysis, stakeholder interaction, and more informed decisions.

9 YOLO and environmental impact

The review identifies YOLO training and retraining as environmentally costly. Its energy, water, and carbon demands motivate more sustainable AI development practices.

  • Environmental impact: YOLO training and retraining consume substantial energy and water and produce significant carbon dioxide emissions.The review links these impacts to concerns about the sustainability of AI development.
  • Environmental impact: The environmental impact of large-scale YOLO training underscores the need for more efficient practices.

10 Conclusion

The review chronicles YOLO’s evolution from YOLOv1 through YOLOv12 and related variants, then projects a shift toward multimodal, context-aware systems. It presents this trajectory as a prospective blueprint for future AI-driven applications.

  • Conclusion: The review retrospectively examines YOLOv1 through YOLOv12 and alternative versions including YOLO-NAS, YOLO-X, YOLO-R, DAMO-YOLO, and Gold-YOLO.
  • Future trajectory: Future YOLO development is anticipated to emphasize multimodal data processing using large language models and natural language processing.
  • Future trajectory: Generative AI and multimodal large language models are expected to broaden YOLO’s utility for sophisticated, context-aware applications.
  • Support: The study received support from the National Science Foundation and the United States Department of Agriculture through specified programs and grants.
  • Study information: No datasets were generated or analyzed during the study.
  • Study information: The authors declare no conflict of interest.
  • Study information: The article is distributed under a Creative Commons Attribution 4.0 International License.

Authors and Affiliations

The paper is authored by a multidisciplinary team affiliated with universities, research centers, and companies across the United States, Ecuador, India, Hong Kong, and China.

  • Academic affiliations: The authors represent expertise spanning biological and environmental engineering, electrical and telecommunications engineering, computer vision, biosystems engineering, soil sciences, and data science.Affiliations include Cornell University, Universidad de las Fuerzas Armadas ESPE, the University of Central Florida, the University of Tennessee, and IISER Thiruvananthapuram.
  • Geographic distribution: The affiliations span the United States, Ecuador, India, Hong Kong, and China.The listed institutions are located in Ithaca, Sangolquí, Orlando, Knoxville, Wadley, Thiruvananthapuram, Billings, College Station, Hong Kong, Kowloon, Hong Kong Science Park, Prosser, and Pullman.
  • Industry and research affiliations: The author team also includes researchers affiliated with Cooper Machine Company, ZenoRobotics, LLC, Washington State University, and the University of Tennessee.These organizations contribute industry and university affiliations spanning machinery, robotics, biological systems engineering, and biosystems engineering.
  • Academic affiliations: Additional academic affiliations include Texas A&M University, The Hong Kong Polytechnic University, City University of Hong Kong, and the Center for Intelligent Multidimensional Data Analysis.These affiliations extend across biological and agricultural engineering, professional education, electrical engineering, and multidimensional data analysis.
Loading 2406.19407v8…