PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLOv1 to YOLOv10: The fastest and most accurate real-time object detection systems

Chien-Yao Wang, Hong-Yuan Mark Liao

arXiv:2408.09332v1cs.CV

TL;DR

Real-time object detection requires methods that balance accuracy, speed, adaptability, and broad applicability. This review re-examines YOLOv1 through YOLOv10, analyzes their technical evolution and influence, and surveys their use across computer vision fields. It concludes that YOLO methods have shaped contemporary real-time computer vision and offer guidance for developing future real-time systems.

  • Problem

    Real-time object detection must serve diverse edge and computer-vision applications, while the YOLO family’s evolution and broader influence require a modern technical review.

  • Method

    The paper reviews YOLOv1–YOLOv10, examines their methods and advantages and disadvantages, analyzes their influence, and surveys applications across computer vision fields.

  • Results

    The review describes YOLO as a unified one-stage detector whose successive methods influenced real-time computer vision through improvements in accuracy, speed, ease of use, and versatility.

  • Takeaways & Limitations

    The review provides readers with guidance for understanding YOLO’s development, applying it to different tasks, and considering future real-time computer-vision directions.

Abstract

from arXiv · show

This is a comprehensive review of the YOLO series of systems. Different from previous literature surveys, this review article re-examines the characteristics of the YOLO series from the latest technical point of view. At the same time, we also analyzed how the YOLO series continued to influence and promote real-time computer vision-related research and led to the subsequent development of computer vision and language models.We take a closer look at how the methods proposed by the YOLO series in the past ten years have affected the development of subsequent technologies and show the applications of YOLO in various fields. We hope this article can play a good guiding role in subsequent real-time computer vision development.

I. INTRODUCTION

Object detection supports many downstream and real-world applications, making real-time edge detection increasingly important. This review traces how YOLO advanced one-stage detection and influenced contemporary computer vision.

  • Real-time object detection on edge devices supports applications including autonomous driving, industrial robots, authentication, healthcare, and surveillance.
  • R-CNN established a breakthrough deep-learning approach by separating object proposal generation from proposal classification.
  • YOLO introduced one-step per-grid prediction, moving object detection toward real-time operation and influencing later one-stage systems.
  • Scaled-YOLOv4 enabled one-stage detection accuracy to surpass contemporary two-stage methods in general object detection.
  • The review covers YOLO methods, their influence on computer vision, and applications across detection, segmentation, pose estimation, and related fields.

A. YOLO (YOLOv1)

YOLOv1 performs object detection in one stage by extracting global image features and making predictions for each grid. Its design also uses IoU-aware confidence and global context to improve prediction quality.

  • YOLOv1 extracts image features with a CNN, obtains global features through fully connected layers, and reshapes them for per-grid prediction.
  • YOLOv1 directly classifies each feature-map grid while predicting bounding-box information in a single-stage design.
  • YOLOv1 uses predicted IoU as a soft objectness label, with final box confidence formed from objectness and classification probability.
  • Fully connected layers provide global context so each grid can use more than local features during prediction.
  • YOLOv2 combines high- and low-resolution features and uses anchor-based prediction within a fully convolutional architecture.

C. YOLOv3

YOLOv3 integrates multi-scale prediction, residual features, and revised label assignment into a one-stage detector, while Gaussian YOLOv3 adds distribution-based bounding-box regression to reduce false positives.

  • YOLOv3: YOLOv3 combines FPN for multi-scale prediction with residual architecture and the DarkNet53 backbone.Its label-assignment design assigns each ground truth to one anchor and changes the assignment scope.
  • YOLOv3: YOLOv3 changes label assignment by assigning each ground truth to one anchor and using features from local to global scales.The design lets each grid obtain multiple-resolution features.
  • Gaussian YOLOv3: Gaussian YOLOv3 converts bounding-box regression into distribution prediction for (x, y, w, h).The Gaussian distribution-based module is intended to reduce false positives.
  • YOLOv4: YOLOv4 replaces FPN with PAN and introduces CSPNet as its backbone while combining techniques aimed at real-time accuracy.The architecture was implemented in both DarkNet and PyTorch YOLOv3 contexts.
  • YOLOv4: YOLOv4 adds receptive-field, attention, activation, and normalization methods that modestly affect inference time while improving accuracy.Its grid-sensitive decoder addresses inaccurate predictions near grid boundaries by keeping target values within an effective gradient range.
  • YOLOv4: YOLOv4 introduces self-adversarial sample generation and memory sharing between GPU and CPU during training.Memory sharing removes the GPU-memory limit on the trained batch size described in the passage.

F. Scaled-YOLOv4

Scaled-YOLOv4 extends YOLOv4 toward edge-to-cloud deployment through architecture scaling and train-from-scratch training, while YOLOv5 continues the PyTorch YOLOv3 design with multiple model specifications.

  • Scaled-YOLOv4: Scaled-YOLOv4 targets both edge and cloud deployment and can obtain high-quality detection results by training from scratch without ImageNet pre-training.Its architecture introduces CSPNet into PAN to improve speed, accuracy, parameter count, and computation.
  • Scaled-YOLOv4: Scaled-YOLOv4’s model scaling lets users obtain the best accuracy for different input resolutions during inference without retraining.Users remove the outputs of corresponding structures to adapt the model.
  • YOLOv5: YOLOv5 continues PyTorch YOLOv3’s design and uses EfficientDet’s scaling pattern to provide models with different specifications.The initial version has an architecture similar to YOLOv3, and the series includes about ten versions.
  • YOLOv5: YOLOv5 adds neighboring grids as positive samples to address recall deficiency.It also enlarges the YOLOv4 center-point decoder’s sigmoid scaling coefficient so neighboring grids can predict the center correctly.
  • Architecture comparison: The referenced architecture figure compares PP-YOLOE, YOLOv6 2.0, YOLOv8, and YOLO-NAS.The passage identifies the systems shown but does not state a performance comparison.

H. PP-YOLO

The PP-YOLO series progresses from YOLOv3-based enhancements to neural architecture search, reparameterization, and distribution-based regression, while YOLOR contributes implicit knowledge modeling and multi-task designs.

  • PP-YOLO: The PP-YOLO series contains PP-YOLO, PP-YOLOv2, PP-PicoDet, and PP-YOLOE.PP-YOLO is based on YOLOv3, while PP-YOLOv2 adds scaled-YOLOv4’s CSPPAN and related mechanisms.
  • PP-PicoDet: PP-PicoDet uses neural architecture search for a mobile-device backbone and introduces YOLOX’s anchor-free decoder.Its search combines ShuffleNetv2 and GhostNet for one-shot architecture search.
  • PP-YOLOE: PP-YOLOE modifies RepVGG into CSPRepResStage and uses TOOD’s distribution-based bounding-box regression.This marks a major architectural change within the PP-YOLO series.
  • Architecture: The referenced architecture figure identifies YOLOR as the system shown.The supplied figure text does not provide an architectural comparison or performance outcome.
  • YOLOR: YOLOR uses a latent variable model as an implicit knowledge encoder and can improve detection effects across YOLO-series models.Its later designs include feature alignment, prediction refinement, and multi-task learning capabilities.
  • YOLOR: YOLOR provides multi-task models for object detection, image classification, and multiple-object tracking, plus pose-estimation models.The passage also identifies advanced AutoML-based training technology as a YOLOR contribution.

J. YOLOX

YOLOX combines CSPNet, an anchor-free head, and dynamic label assignment, while later YOLO systems extend related ideas through quantization-focused design, auxiliary training, and efficient architectures.

  • YOLOX: YOLOX combines CSPNet with FCOS’s anchor-free head and replaces manual label assignment with the SimOTA dynamic assignment method.Its design also improves OTA for practical use.
  • YOLOv6: YOLOv6 uses RepVGG initially, later introduces CSPNet, and emphasizes quantization stability through RepOPT, QAT, and knowledge distillation.These methods are described as enhancing the accuracy of quantized models.
  • YOLOv7: YOLOv7 introduces trainable auxiliary architectures removable or integrable during inference and uses ELAN and E-ELAN for its architecture.It also proposes coarse-to-fine label assignment and auxiliary losses.
  • YOLOv7: YOLOv7’s partial auxiliary loss updates selected features with auxiliary information while leaving other features focused on the target task.The developers report improvement on the main tasks.
  • YOLOv7: YOLOv7 provides object-detection, instance-segmentation, and pose-estimation models with reported real-time state-of-the-art performance in these tasks.The passage presents this as a set of task-related models from YOLOv7.
  • Architecture: The referenced architecture figure identifies DAMO-YOLO as one of the systems shown.The supplied figure text does not state a comparison or outcome.

M. DAMO-YOLO

The review presents DAMO-YOLO, YOLOv8, YOLO-NAS, and Gold-YOLO as systems that improve architecture efficiency, training, deployment interfaces, or feature integration. Their designs include neural architecture search, multi-stage training, downstream-task APIs, and transformer-based global feature aggregation.

  • DAMO-YOLO: DAMO-YOLO uses MAE-NAS to search CSPNet and ELAN for a more efficient architecture.It also modifies feature integration, the prediction head, and label assignment.
  • YOLOv8: YOLOv8 refactors YOLOv5, optimizes underlying code, and provides APIs connecting detection with downstream tasks.The review describes it as a technology integration platform rather than an entirely new YOLO version.
  • YOLO-NAS: YOLO-NAS uses AutoNAC NAS to design a quantization-friendly architecture and applies multi-stage training with pre-training, pseudo-labeling, knowledge distillation, and DFL.The cited passage does not provide further technical details about the architecture.
  • Architectural overview: The reviewed architectures are illustrated through diagrams for Gold-YOLO and YOLOv9.The supplied figure captions identify the Gold-YOLO architecture in Figure 12 and the YOLOv9 architecture in Figure 13.
  • Gold-YOLO: Gold-YOLO replaces PAN with a Gather-and-Distribute mechanism that integrates layer features into global features using transformers.Its overall architecture is described as similar to YOLOv6 3.0, with masked image modeling used during training.

Q. YOLOv9

YOLOv9 introduces Programmable Gradient Information to preserve and share task-relevant information, while YOLOv10 combines transformer-based global feature extraction with dual label assignment to remove post-processing. These designs are presented within the YOLO series’ emphasis on simple, deployable real-time detection systems.

  • YOLOv9: YOLOv9 introduces Programmable Gradient Information (PGI) to enhance model interpretability, robustness, and versatility.PGI uses reversible architecture and multi-level information to retain original data and task-relevant information.
  • YOLOv9: PGI shares information retained by a reversible auxiliary branch with the main branch to address the information bottleneck in deep neural networks.The cited description distinguishes this use from general-purpose reversible architectures that simply maximize retained information.
  • YOLOv10: YOLOv10 adds a transformer-based module for global feature extraction and uses one-to-many and one-to-one matching in a dual head.The review states that this enables end-to-end detection without DETR-style post-processing.
  • Series characteristics: The YOLO series is characterized by relatively simple frameworks and relatively easy deployment.These characteristics are introduced as recurring properties of the series.
  • Series characteristics: YOLO’s one-stage detection approach introduced a simpler framework that influenced subsequent computer vision research.The review relates this approach to earlier DeepMultiBox and OverFeat research.

B. Better

The review attributes YOLO’s practical strengths to training methods compatible with current techniques, hardware-oriented architecture design, adaptability across datasets and domains, and use in multiple computer vision tasks. These properties support deployment from edge devices to broader application settings.

  • Better Training Technique: YOLO training techniques are described as advanced and complementary to current training technologies.The review notes that many studies validate methods on foundation models without examining compatibility with the current state of the art.
  • Faster Architecture: YOLO architectures target actual hardware inference speed rather than relying only on computational-load reductions.The review discusses device-specific criteria spanning edge to cloud and cites design responses including Fully CSPOSANet and CSPDark-Net.
  • Faster Architecture: YOLOv6 uses RepVGG as an efficient backbone, while DAMO-YOLO uses NAS to search efficient architectures in CSPNet and ELAN.These examples illustrate architecture choices made to improve inference efficiency.
  • Stronger Adaptability: YOLO systems can be trained across domains and datasets without relying on many domain-specific pre-trained models.The review also cites evolutionary hyperparameter search and anchor-free designs as supporting broader application.
  • Stronger Capability: YOLO-based models have achieved strong performance across multiple computer vision tasks and application domains.The supplied passage introduces this broader capability after describing the series’ adaptability.
  • Stronger Capability: The YOLO series has been widely used in fields beyond object detection, including instance segmentation, pose estimation, and 3D object detection.The review frames these uses as representative real-time computer vision applications.

A. Multiple Object Tracking

YOLO-based systems extend real-time processing to multiple object tracking, instance segmentation, and self-driving perception. The reviewed methods reduce computation by sharing detection and tracking features, decomposing segmentation outputs, or adapting YOLO architectures to vehicle-perception tasks.

  • Multiple Object Tracking: ROLO directly uses YOLO detections with LSTM-based processing for single-object tracking and extends this design to multiple-object tracking.The review contrasts this with methods that crop detected regions and extract tracking features through additional networks.
  • Multiple Object Tracking: JDE outputs re-identification features within the detection pipeline to support tracking.The cited passage introduces JDE as another YOLO-related tracking approach.
  • Instance Segmentation: YOLACT and YOLACT++ decompose instance segmentation into prototypes and coefficients, reducing the operations required for prediction.The output masks are formed by combining predicted coefficients with prototypes.
  • Instance Segmentation: Insta-YOLO and Poly YOLO predict instance-segmentation results using polygon representations of binary masks.The lower-dimensional representation reduces computation but can introduce distortion.
  • Self-Driving Perception: YOLOP and YOLOPv2 use CSPNet and ELAN as main architectures for object detection, area detection, and lane prediction.These systems are presented as visual-perception methods for self-driving scenarios.
  • Self-Driving Perception: HybridNet, YOLOPv3, and YOLOPX modify different YOLO versions to perform self-driving tasks.The review groups these systems with YOLOP and YOLOPv2 as real-time vehicle-perception approaches.

D. Human Pose Estimation

The YOLO series has been adapted to human pose estimation and extended across 3D, video, and face-detection applications. These variants preserve YOLO’s role in real-time computer vision systems.

  • Human Pose Estimation: YOLO-based pose methods add keypoint prediction to object detection, using either separate pose/keypoint expressions or grid-relative regression.KAPAO separates human pose and keypoint objects, while YOLO-Pose predicts keypoints relative to grid centers.
  • 3D Object Detection: YOLO variants generalize 2D detection to 3D using images, RGB-D images, or combined image-and-LIDAR inputs.The cited systems include ComplexYOLO, Expandable YOLO, YOLO 6D, and YOLO 3D.
  • Video Object Detection: YOLO-based systems support video object detection and streaming perception.YOLOV and YOLOV++ address video object detection, while Stream YOLO targets streaming perception.
  • Face Detection: YOLO-derived models are also applied to face detection, which the review identifies as a prominent object-detection subfield.The examples include YOLO-Face, YOLO-Face v2, and YOLO5Face.

H. Image Segmentation

YOLO has been incorporated into image segmentation and open-vocabulary systems while remaining compatible with diverse neural architectures. The review presents this adaptability as part of YOLO’s broader versatility across real-time vision tasks.

  • Image Segmentation: Fast-SAM combines YOLO with SAM for general image segmentation and can greatly improve inference speed.The cited combination uses YOLO’s real-time characteristics alongside the Segment Anything Model.
  • Open-Vocabulary Detection: YOLO-based systems combine with CLIP to perform 2D and 3D open-vocabulary object detection.YOLO-World addresses 2D open-vocabulary detection, while Open-YOLO 3D addresses the 3D setting.
  • Neural Architectures: YOLO is compatible with transformer, state-space, spiking, graph, and KAN neural-network architectures.The review lists ViT, MAMBA, SNN, GNN, GCN, and KAN-based YOLO variants.
  • Review Scope: The review examines YOLO’s evolution, contributions, accuracy, speed, ease of use, and versatility across application fields.It aims to help readers understand real-time computer-vision development, applicable tasks, and possible future directions.
Loading 2408.09332v1…