PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception

Mengqi Lei, Siqi Li, Yihong Wu, Han Hu, You Zhou, Xinhu Zheng, Guiguang Ding, Shaoyi Du, Zongze Wu, Yue Gao

arXiv:2506.17733v2cs.CV

TL;DR

Existing YOLO models are limited in their ability to capture global multi-to-multi high-order correlations, which constrains detection in complex scenarios. YOLOv13 addresses this with adaptive hypergraph-based correlation enhancement, FullPAD feature distribution, and lightweight convolutional blocks, achieving improved MS COCO performance with fewer parameters and FLOPs.

  • Problem

    Existing YOLO models are restricted to local pairwise correlations and lack global multi-to-multi high-order correlation modeling for complex scenarios.

    Source: §I · p. 2, A. Evolution of YOLO Detectors · p. 3

  • Method

    YOLOv13 combines HyperACE for adaptive high-order correlation modeling, FullPAD for full-pipeline feature aggregation and distribution, and lightweight blocks based on depthwise separable convolutions.

    Source: §I · p. 2, §V · p. 10

  • Results

    YOLOv13 achieves state-of-the-art detection performance while remaining lightweight on MS COCO, with YOLOv13-N improving mAP by 3.0% over YOLO11-N and 1.5% over YOLOv12-N.

    Source: §I · p. 2, §V · p. 10

  • Takeaways & Limitations

    Adaptive high-order correlation modeling supports accurate detection while the proposed design maintains low parameter and computational costs.

    Source: §I · p. 2, B. Comparison with Other Methods · p. 8, §V · p. 10

Abstract

from arXiv · show

The YOLO series models reign supreme in real-time object detection due to their superior accuracy and computational efficiency. However, both the convolutional architectures of YOLO11 and earlier versions and the area-based self-attention mechanism introduced in YOLOv12 are limited to local information aggregation and pairwise correlation modeling, lacking the capability to capture global multi-to-multi high-order correlations, which limits detection performance in complex scenarios. In this paper, we propose YOLOv13, an accurate and lightweight object detector. To address the above-mentioned challenges, we propose a Hypergraph-based Adaptive Correlation Enhancement (HyperACE) mechanism that adaptively exploits latent high-order correlations and overcomes the limitation of previous methods that are restricted to pairwise correlation modeling based on hypergraph computation, achieving efficient global cross-location and cross-scale feature fusion and enhancement. Subsequently, we propose a Full-Pipeline Aggregation-and-Distribution (FullPAD) paradigm based on HyperACE, which effectively achieves fine-grained information flow and representation synergy within the entire network by distributing correlation-enhanced features to the full pipeline. Finally, we propose to leverage depthwise separable convolutions to replace vanilla large-kernel convolutions, and design a series of blocks that significantly reduce parameters and computational complexity without sacrificing performance. We conduct extensive experiments on the widely used MS COCO benchmark, and the experimental results demonstrate that our method achieves state-of-the-art performance with fewer parameters and FLOPs. Specifically, our YOLOv13-N improves mAP by 3.0\% over YOLO11-N and by 1.5\% over YOLOv12-N. The code and models of our YOLOv13 model are available at: https://github.com/iMoonLab/yolov13.

I. INTRODUCTION

YOLOv13 addresses the limitation of existing YOLO architectures, which primarily model local or pairwise correlations, by introducing adaptive high-order correlation modeling and lightweight feature-processing blocks. Experiments on MS COCO report improved detection performance while maintaining a lightweight design.

  • Existing YOLO architectures are limited to local pairwise correlation modeling, restricting detection performance in complex scenarios.Source: §I · pp. 1–2, A. Evolution of YOLO Detectors · p. 3
  • YOLOv13-N/S improve mAP by 1.5%/0.9% over YOLOv12-N/S and by 3.0%/2.2% over YOLO11-N/S, respectively.Source: §I · p. 2
  • YOLOv13 uses adaptive hypergraphs to explore latent high-order correlations for accurate and robust object detection.Source: §I · p. 2
  • HyperACE captures high-order correlations, FullPAD distributes enhanced features throughout the pipeline, and lightweight blocks reduce parameters and computational complexity.Source: §I · p. 2
  • YOLOv13 achieves state-of-the-art detection performance while remaining lightweight on the MS COCO benchmark.Source: §I · p. 2

B. High-Order Correlation Modeling

The paper frames global multi-to-multi high-order correlation modeling as a missing capability in current YOLO detectors. It introduces an adaptive hypergraph mechanism and FullPAD to improve cross-location, cross-scale feature interaction and information flow.

1) Adaptive Hypergraph Computation:

The adaptive hypergraph computation paradigm represents visual features with continuously learned vertex participation in dynamically generated hyperedges. Hypergraph convolution then aggregates features through hyperedges and disseminates enhanced representations back to vertices.

Adaptive Hypergraph Computation:

2) C3AH for Adaptive High-Order Correlation Modeling:

C3AH combines CSP-style branch splitting with adaptive hypergraph computation to capture global high-order semantic interactions across spatial positions. It projects input features, processes one branch for correlation enhancement, and fuses the result with a lateral connection.

C3AH design

  • C3AH integrates an adaptive hypergraph computation module into a CSP bottleneck branch-split structure for global high-order semantic aggregation.The module captures high-order interactions across spatial positions.Source: 2) C3AH for Adaptive High-Order… · p. 5
  • The input feature map is projected through two 1 × 1 convolutions into a shared hidden dimension before correlation processing.The projected features are represented as X and Xlateral.Source: 2) C3AH for Adaptive High-Order… · p. 5

Adaptive correlation computation

  • The projected feature X is flattened into vertex features and sent to adaptive hypergraph computation to obtain correlation-enhanced features.Xlateral serves as the lateral-connection branch.Source: 2) C3AH for Adaptive High-Order… · p. 5

Feature fusion

3) Structure of HyperACE:

HyperACE fuses multi-scale backbone features and divides them into complementary global, local, and shortcut branches. Their outputs are recombined to provide multi-level visual correlation perception across global-local and high-low orders.

Multi-scale input

  • HyperACE resizes B3 and B5 to B4’s spatial size and applies a 1 × 1 convolution to obtain fused feature Xb.The inputs are the last three backbone stages.Source: 3) Structure of HyperACE · p. 5

Three-branch structure

Output fusion

  • HyperACE concatenates the three branch outputs and fuses them with a 1 × 1 convolution, preserving complementary global-local and high-low-order perception.Source: 3) Structure of HyperACE · p. 5

C. Full-Pipeline Aggregation-and-Distribution Paradigm

FullPAD redistributes HyperACE’s correlation-enhanced features across the network rather than restricting information flow to the conventional backbone–neck–head sequence. It uses resized, channel-adjusted features and gated fusion to support fine-grained information flow and representation synergy.

FullPAD paradigm

  • FullPAD collects multi-scale backbone features, processes them with HyperACE, and redistributes the enhanced features through tunnels to locations across the pipeline.The design is intended to improve gradient propagation and detection performance.Source: C. Full-Pipeline… · p. 6

Feature redistribution

  • The correlation-enhanced feature Y is resized to each stage’s spatial resolution and channel-adjusted with a 1 × 1 convolution.This produces stage-specific features for i ∈ {3, 4, 5}.Source: C. Full-Pipeline… · p. 6
  • For an arbitrary stage feature map Fi, FullPAD applies gated fusion to combine the original feature with the correlation-enhanced feature.The fusion is applied at each stage receiving redistributed information.Source: C. Full-Pipeline… · p. 6

Adaptive fusion

  • A learnable scalar γ adaptively balances the correlation-enhanced feature and the original feature.FullPAD transmits enhanced features to seven destinations through three tunnels.Source: C. Full-Pipeline… · p. 6

Full-pipeline integration

  • FullPAD integrates correlation-enhanced features into different pipeline stages, enabling visual-correlation information to support perception of complex scenarios.Source: C. Full-Pipeline… · p. 6

D. Model Lightweighting with Depth-Separable Convolution

YOLOv13 lightweightens feature extraction with depthwise-separable large-kernel convolutions and DS-series blocks. These blocks preserve performance while reducing computational cost, with deployment across the backbone, neck, and HyperACE’s local branch.

DS-series design

Efficiency results

  • Up to 30% parameter reduction and up to 28% GFLOPs reduction are achieved across YOLOv13 model sizes using DS-series blocks.DS-C3k2 is used in the Backbone and Neck, while DS-C3k serves as HyperACE’s low-order extractor.Source: D. Model Lightweighting with… · p. 6

IV. EXPERIMENTS

The experiments evaluate YOLOv13 against real-time detectors on MS COCO and additionally assess cross-domain generalization on Pascal VOC 2007.

Dataset

  • The study compares YOLOv13 with existing real-time object detectors using the MS COCO benchmark.MS COCO is used to evaluate the proposed model and other state-of-the-art detectors.Source: §IV · p. 6, 1) Dataset · pp. 6–7, Table I · p. 7
  • The evaluation includes cross-domain testing by training models on MS COCO and testing them on Pascal VOC 2007.Pascal VOC 2007 serves as a supplementary benchmark for distribution-shift generalization.Source: 1) Dataset · p. 7

2) Implementation Details:

The implementation follows controlled training settings and evaluates YOLOv13 against prior detectors using quantitative, qualitative, and generalization comparisons.

Implementation Details

  • All YOLOv13 variants are trained for 600 epochs with batch size 256, SGD, 640 × 640 inputs, and standard YOLO augmentations.The variants use 4, 8, 8, and 12 hyperedges for Nano, Small, Large, and Extra-Large models, respectively.Source: 2) Implementation Details · p. 7

Comparison with Other Methods

  • YOLOv13 achieves state-of-the-art MS COCO performance across variants while remaining lightweight.The comparison uses the same GPUs for YOLOv13, YOLO11, and YOLOv12, with other methods trained using official settings.Source: B. Comparison with Other Methods · p. 7
  • 1.5%, 0.9%, 0.4%, and 0.4% APval 50:95 improvements over YOLOv12 are reported for Nano, Small, Large, and Extra-Large models.The corresponding APval 50 improvements are 1.8%, 1.0%, 0.9%, and 0.9%.Source: B. Comparison with Other Methods · p. 8
  • 0.1% APval 50:95 improvement over RT-DETRv2-R18 is reported for YOLOv13-S, alongside 55.0% fewer parameters and 65.3% fewer FLOPs.The comparison demonstrates an accuracy and efficiency advantage against a ViT-based detector.Source: B. Comparison with Other Methods · p. 8
  • YOLOv13 produces more accurate detections in complex scenes, including small objects, occluded plants, and tennis rackets missed or misclassified by prior YOLO models.The qualitative comparisons cover YOLOv10, YOLO11, YOLOv12, and YOLOv13 on MS COCO.Source: B. Comparison with Other Methods · p. 8, Figure 5 · p. 8
  • YOLOv13 shows satisfactory generalization when trained on MS COCO and tested on Pascal VOC 2007.The reported gains over previous models include 1.0% and 0.4% for Nano and Small models.Source: B. Comparison with Other Methods · pp. 8–9, Table II · p. 8

C. Ablation Study

Ablation studies examine FullPAD distribution, adaptive hypergraph visualization, training duration, and inference latency to characterize YOLOv13’s design and deployment behavior.

FullPAD and HyperACE

  • Removing HyperACE decreases APval 50:95 and APval 50 by 0.9% and 1.1%, respectively, demonstrating the effectiveness of adaptive correlation enhancement.The ablation uses YOLOv13-Small with FullPAD distribution settings.Source: 1) FullPAD and HyperACE · p. 9
  • 0.2%, 0.4%, and 0.3% APval 50:95 decreases result when enhanced features are distributed only to the backbone-neck, in-neck, or neck-head locations.These reductions are measured relative to the full model.Source: 1) FullPAD and HyperACE · p. 9

Hyperedge Visualization

  • Adaptive hyperedges model high-order correlations among multiple foreground objects and between foreground targets and background scenes.Examples include skis with ski sticks, cars with traffic lights, rackets with courts, and gloves with fields.Source: 1) FullPAD and HyperACE · p. 9, Figure 6 · p. 9

2) Number of hyperedges:

The study balances hyperedge count, efficiency, training behavior, and deployment speed while preserving YOLOv13’s accuracy through its lightweight design.

Number of Hyperedges

  • Fewer hyperedges reduce parameters and computational effort but lower performance because scene correlations are modeled insufficiently.Increasing the count to 16 continues improving detection performance while adding parameters and computational costs.Source: 2) Number of hyperedges · pp. 9–10
  • YOLOv13 uses 4, 8, 8, and 12 hyperedges for Nano, Small, Large, and Extra-Large variants to balance performance and computational complexity.These settings are selected after evaluating different hyperedge counts.Source: 2) Number of hyperedges · p. 10

Lightweight Blocks

  • Replacing DS-series blocks with vanilla convolutions causes only a 0.1% APval 50 decrease and no APval 50:95 decrease.The DS-series blocks reduce FLOPs by 1.1 G and 4.2 G in the evaluated Nano and Small models.Source: 3) DS blocks · p. 10

Training Epochs

  • Training for 600 epochs achieves 41.6% and 48.0% APval 50:95 for YOLOv13 Nano and Small, while additional epochs degrade performance through overfitting.The result identifies the selected training duration for these variants.Source: 4) Training epochs · p. 10

Conclusion

  • YOLOv13 combines adaptive global high-order correlation modeling, FullPAD feature distribution, and depthwise-separable-convolution blocks to maintain accuracy with lower parameters and FLOPs.The conclusion presents these components as the basis of the real-time detector.Source: §V · p. 10

Related papers

Loading 2506.17733v2…