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

TL;DR

This paper provides a comprehensive, decade-long review of the You Only Look Once (YOLO) object detection series, tracing its evolution from YOLOv1 to the latest YOLOv12. It highlights the continuous advancements in speed, accuracy, and computational efficiency, and discusses YOLO's transformative impact across critical application areas like autonomous vehicles, healthcare, and manufacturing.

What Problem Does This Paper Address?

Object detection is a fundamental task in computer vision, enabling machines to identify and locate objects in images and videos. However, real-world scenarios present significant challenges: varying object scales, orientations, lighting conditions, and occlusions. Furthermore, many applications demand real-time performance, requiring a delicate balance between accuracy and speed. This paper addresses the need for a comprehensive understanding of how the YOLO family of object detectors has evolved over the past decade to tackle these challenges, offering a unified view of its progress and impact.

Methods

The paper employs a reverse chronological analysis to review the YOLO series, starting from the most recent versions (YOLOv12, YOLO11) and moving backward to the foundational YOLOv1. It details the architectural innovations, new techniques, and optimizations introduced in each version. Key advancements include:

  • Unified Architecture: YOLO's core innovation was combining region proposal and classification into a single neural network, enabling end-to-end learning and significantly reducing computation time.
  • Backbone Evolution: The series has seen continuous development in backbone architectures (e.g., Darknet-19, Darknet-53, CSPDarknet-53, GELAN) to improve feature extraction.
  • Attention Mechanisms: More recent versions (YOLOv12) integrate attention modules (e.g., Area Attention, C2PSA) to enhance feature processing and capture long-range dependencies.
  • Anchor-Free Approaches: Some variants have moved towards anchor-free designs, simplifying the process and improving efficiency.
  • Advanced Training Techniques: Innovations like mosaic data augmentation, CutMix, Mish activation, and sophisticated label assignment strategies have been incorporated.
  • Multi-Scale Detection: Many versions utilize multi-scale feature maps and prediction heads to effectively detect objects of various sizes.

The paper also reviews alternative YOLO-derived versions like YOLO-NAS, YOLO-X, YOLO-R, DAMO-YOLO, and Gold-YOLO, which leverage techniques like Neural Architecture Search (NAS) and novel fusion mechanisms.

Datasets

The paper mentions the use of standard datasets for training and evaluation, including:

  • MS COCO: A widely used benchmark dataset for object detection, segmentation, and captioning.
  • Objects365: Another large-scale dataset for object detection.
  • Roboflow 100: A dataset used for training YOLO-NAS.
  • TT100K: A dataset specifically for traffic sign detection.
  • SODA-A: A dataset used for aircraft detection.
  • GRAZPEDWRI-DX: A dataset for pediatric wrist trauma X-ray images.
  • LUNA 16: A dataset for lung cancer detection.
  • German traffic sign recognition standard: Used for traffic sign detection.

No new datasets were introduced in this review paper.

Key Insights and Ideas

  • Decadal Progress: The YOLO series has consistently pushed the boundaries of real-time object detection, demonstrating a remarkable decade of innovation.
  • Trade-offs are Key: Each YOLO version represents a careful balance between speed, accuracy, and computational cost, with different variants optimized for specific use cases (e.g., edge devices vs. high-performance servers).
  • Architecture Matters: The evolution of YOLO is largely driven by advancements in its backbone, neck, and head architectures, incorporating techniques like CSPNets, attention mechanisms, and decoupled heads.
  • Beyond Core Detection: Newer YOLO versions are expanding their capabilities beyond basic object detection to include instance segmentation, pose estimation, and oriented object detection.
  • The Rise of Attention: Attention mechanisms are becoming increasingly crucial in YOLO models, enabling them to focus on relevant features and improve performance in complex scenarios.
  • AGI Integration: The future of YOLO likely involves integration with Large Language Models (LLMs) and other AI systems to achieve more sophisticated, context-aware, and human-like intelligence.

Key Results

The paper highlights significant improvements across YOLO versions:

  • YOLOv12: Achieves state-of-the-art performance with competitive inference speeds, demonstrating a substantial reduction in latency (e.g., 1.64 ms on a T4 GPU for YOLOv12-N) and improved accuracy (up to 55.2% mAP for YOLOv12-X on MS COCO). It also shows superior object contour definition.
  • YOLOv11: Offers improved computational efficiency and accuracy, with faster inference times than YOLOv10.
  • YOLOv10: Achieves unprecedented performance with significantly reduced latency due to the elimination of NMS, offering variants with very low latencies (e.g., 1.84 ms for YOLOv10-N).
  • YOLOv9: Introduces Programmable Gradient Information (PGI) and GELAN architecture, enhancing robustness and generalization, with significant parameter and computational load reductions compared to YOLOv8.
  • YOLOv8: Demonstrates significant improvements in object detection, segmentation, and other tasks, with high accuracy and speed (e.g., 280 FPS on an NVIDIA A100 for YOLOv8x).
  • Overall Impact: YOLO models have significantly impacted diverse fields, enabling real-time applications in autonomous vehicles, healthcare, manufacturing, security, and agriculture, often outperforming previous state-of-the-art models.

Limitations and Future Work

Limitations:

  • Small and Rotated Objects: Many YOLO versions, including recent ones like YOLOv11 and YOLOv12, still struggle with detecting small, low-resolution, or rotated objects, especially in crowded or complex environments. This is often attributed to downsampling processes and fixed anchor box limitations.
  • Overfitting: YOLOv11 is noted as being susceptible to overfitting on limited datasets.
  • Computational Efficiency vs. Accuracy Trade-off: While efficiency has improved, a trade-off with accuracy, particularly in complex scenarios or on resource-constrained devices, persists.
  • Limited Adoption/Validation: Some newer versions, like YOLOv10, have not yet seen widespread adoption or extensive real-world validation, making it difficult to fully assess their potential and limitations.
  • Environmental Impact: Training YOLO models is energy-intensive, raising concerns about sustainability and the need for more efficient practices.

Future Work:

  • Edge and IoT Deployment: Enhancing efficiency and accuracy for ultra-low-power devices, exploring model pruning, quantization, and specialized hardware.
  • Embodied AI Integration: Integrating YOLO with physical entities for more natural and human-like interaction with the real world.
  • AGI Integration: Merging YOLO with LLMs and other cognitive functions to create more intelligent agents capable of complex, real-time tasks.
  • Environmental Friendliness: Developing more efficient training methods to reduce energy consumption and carbon footprint.
  • Improved Evaluation Metrics: Developing novel metrics to capture emerging capabilities in edge computing, multi-modal fusion, and adaptive architectures.
  • Addressing Detection Challenges: Future research should focus on improving scale-invariant feature extraction, handling occlusions, and developing more robust models for dynamic and adverse environments.
  • Federated Learning: Exploring federated learning paradigms to enable continuous learning and data privacy.

Summary updated

BibTeX
@article{sapkota2025yolo,
  author = {Ranjan Sapkota and Marco Flores-Calero and Rizwan Qureshi and Chetan Badgujar and Upesh Nepal and Alwin Poulose and Peter Zeno and Uday Bhanu Prakash Vaddevolu and Sheheryar Khan and Maged Shoman and Hong Yan and Manoj Karkee},
  title = {{YOLO advances to its genesis: a decadal and comprehensive review of the You Only Look Once (YOLO) series}},
  journal = {Artif. Intell. Rev.},
  year = {2025},
  doi = {10.1007/s10462-025-11253-3},
}
Figure