YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception YOLOv13, by Tsinghua University, Taiyuan University of Technology, Beijing Institute of Technology, Shenzhen University, The Hong Kong University of Science and Technology (Guangzhou), The Hong Kong University of Science and Technology, Xi’an Jiaotong University 2025 arXiv v1 (
HyperACE: YOLOv13 proposes Hypergraph-based Adaptive Correlation Enhancement (HyperACE) mechanism that adaptively exploits latent high-order correlations, achieving efficient global cross-location and cross-scale feature fusion and enhancement.
FullPAD: Full-Pipeline Aggregation-and-Distribution (FullPAD) paradigm based on HyperACE is designed, which effectively achieves fine-grained information flow and representation synergy within the entire network by distributing correlation-enhanced features to the full pipeline.
DSConv: Finally, depthwise separable convolutions (DSConv) are leveraged that significantly reduce parameters and computational complexity without sacrificing performance.
Outline
YOLOv13: Overall Architecure
YOLOv13: HyperACE
YOLOv13: FullPAD
YOLOv13: DSConv
Experimental Results
1. YOLOv13: Overall Architecure
Press enter or click to view image in full size
YOLOv13: Overall Architecure
YOLOv13 model first extractsmulti-scale feature maps B1, B2, B3, B4, B5 using a backbone network similar to previous works, but in which the large-kernel convolutions are replaced with the proposed lightweight DS-C3k2 blocks.
The method gathers and forwards these features into the proposed HyperACE module, achieving cross-scale cross-position features high-order correlation adaptive modeling and feature enhancement.
Subsequently, FullPAD paradigm leverages three separate tunnels to distribute the correlation-enhanced feature to the connection between the backbone and neck.
Finally, the output feature maps of the neck are forwarded into detection heads to achieve multi-scale object detection.
2. YOLOv13: HyperACE
(Please skip this part for quick read.)
YOLOv13: HyperACE
1) Adaptive Hypergraph Computation: The proposed adaptive hypergraph is defined as G = {V, A}, where V is the vertex set, and A is the adaptive hyperedge set. In an adaptive hypergraph, each hyperedge connects all vertices, where each vertex participates in the hyperedge with a continuous and differentiable contribution degree.
The proposed adaptive hypergraph is represented using a continuous participation matrix A where A_i,m denotes the participation degree ofvertex i in hyperedge m.
The proposed adaptive hypergraph computation paradigm consists of two stages, i.e., adaptive hyperedge generation and hypergraph convolution, as shown in Fig. 3.
2) Adaptive Hyperedge Generation: Specifically, let X denote the vertex features, where C is the number of feature channels. The proposed method first uses global average pooling and max pooling to generate context vectors favg and fmax, respectively. Then, the context vectors are concatenated to obtain global vertex context fctx:
Subsequently, a mapping layer is leveraged to generate the global offset ΔP from the vertex context:
To calculate the participation degree of each vertex, another projection layer is leveraged to generate the vertex query vector zi from vertex feature xi:
To further increase feature diversity, a multi-head mechanism is introduced to split zi into h subspaces {ˆz^τ_i} along the feature dimension, then the similarity between the i-th vertex query vector and the m-th prototype is computed:
Thus, the overall similarity is defined as the average of all subspace similarities.
Finally, the similarity between the vertex query vector and the prototype is used as the contribution of the vertex to the hyperedge, and normalized across vertices to obtain the continuous participation matrix A, formulated as:
3) Hypergraph Convolution: Specifically, each hyperedge first collects features from all vertices and applies a linear projection to form the hyperedge feature. Then, the hyperedge features are disseminated back to the vertices to update their representations:
4) C3AH for Adaptive High-Order Correlation Modeling: C3AH block retains the CSP bottleneck branch-split mechanism as in CSPNet:
Then, X is flattened as the vertex features and is sent to the adaptive hypergraph computation (AHC) module, to obtain correlation-enhanced features:
Finally, X_h and X_lateral are concatenated along the channel dimension and then fused by a 1 × 1 convolutional layer to obtain the output of C3AH block.
5) Structure of HyperACE: HyperACE first takes the feature maps B3, B4, and B5 from the last three stages of the backbone as input, then resizes B3 and B5 to the same spatial size as B4, and aggregates them via a 1×1 convolutional layer to obtain the fused feature X_b.
Subsequently, X_b is split along the channel dimension into three feature groups.X^h_b, X^l_b, and X^s_b are used for global high-order correlation modeling, local low-order correlation modeling, and shortcut connection, respectively.
In the high-order correlation modeling branch, X^h_b is forwarded into K parallel C3AH blocks to explore various latent high-order correlations and obtain enhanced features:
These K enhanced features are then concatenated along the channel dimension to obtain the output of the high-order correlation modeling branch X_h.
In the local low-order correlation modeling branch, L stacked DS-C3k modules is used to capture fine-grained local information:
The shortcut branch directly retains the original visual information, i.e., X^s = X^s_b.
Finally, the outputs of three branches are concatenated along the channel dimension and fused by a 1×1 convolutional layer to obtain the final output of HyperACE:
HyperACE achieves complementary multi-level visual correlation perception across global-local and high-low orders.
3. YOLOv13: FullPAD
FullPAD collects multi-scale feature maps from the backbone and forwards them into HyperACE, and then redistributes the enhanced features to various locations throughout the pipeline via different FullPAD tunnels, as in Figure 2 above.
In practice, after the correlation-enhanced feature Y is obtained from B3, B4, B5 using HyperACE, FullPAD resizes Y to the spatial resolution of each stage and adjusts its channel dimension using a 1 × 1 convolutional layer:
Then, for an arbitrary feature map Fi at stage i, a gated fusion is leveraged to achieve information flow and fusion:
where γ is a learnable scalar parameter.
FullPAD enables the model to fully utilize the visual correlation information to perceive complex scenarios effectively.
2) DS-Bottleneck: In the DS-Bottleneck block, two DSConv blocks are cascaded, among which the first block is a fixed 3 × 3 depthwise-separable convolution, and the second block is a large-kernel (k×k) depthwise-separable convolution:
3) DS-C3k: The DS-C3k block inherits from the standard CSP-C3 structure from YOLOv5. Specifically, the input feature is first forwarded into a 1 × 1 convolutional layer to reduce feature channels, and is then processed by n cascaded DS-Bottleneck blocks. Meanwhile, a lateral 1 × 1 convolutional branch is applied to the input feature. Finally, the features from two branches are concatenated along the channel dimension, and a 1 × 1 convolutional layer is leveraged to restore the feature channels.
This design retains the cross-channel branching of CSP structure while integrating depthwise-separable lightweight bottlenecks.
4) DS-C3k2: is derived from the C3k2 structure in YOLOv11. Specifically, a 1×1 convolutional layer is first applied to unify the channels. Then, the features are split into two parts, with one part fed into multiple DS-C3k modules and the other passed through a shortcut connection. Finally, the outputs are concatenated and fused with a 1×1 convolutional layer.
YOLOv13 model extensively uses DS-C3k2 blocks as the basic feature extraction module in both the Backbone and the Neck. In HyperACE, the DS-C3k block is leveraged as the low-order feature extractor. This design achieves up to 30% parameter reduction and up to 28% GFLOPs reduction across all YOLOv13 model sizes.
5. Experimental Results
Press enter or click to view image in full size
Comparison with Other Methods
All variants of YOLOv13 model achieve state-of-the-art performance while remaining lightweight.
Press enter or click to view image in full size
Qualitative comparison on MS COCO dataset
YOLOv13-N model can accurately detect objects in complex multi-object scenes. In contrast, previous models miss small objects such as bowls and vases. This is due to the fact that our proposed HyperACE can establish high-order correlations between multiple related objects, enabling accurate detection of multiple targets in complex scenes.
YOLOv13 model and previous YOLO models are trained on the MS COCO dataset and tested on the Pascal VOC 2007 dataset.
YOLOv13 model achieves satisfactory generalization performance.
For the Nano model of YOLOv13, the inference latency of 1.25 ms and 1.97 ms are achieved on the RTX 4090 and the Tesla T4 GPU, respectively. Considering the deployment condition without a GPU, the Nano model can also achieve an inference speed of 25 FPS (39.97 ms) on a CPU (Intel Xeon Platinum 8352V).
For the Small model with better performance, the inference latency on the Tesla T4 GPU is still less than 3 ms.
For the Extra-Large model of YOLOv13, the inference latency on the Tesla T4 GPU is 14.67 ms, while on the 4090 GPU it is only 3.1 ms.
(For ablation studies, please read the paper directly.)
