In a real-world traffic scenario, varying-scale objects are usually distributed in a cluttered background, which poses great challenges to accurate detection. Although current Mamba-based methods can efficiently model long-range dependencies, they still struggle to capture small objects with abundant local details, which hinders joint modeling of local structures and global semantics. Moreover, state-space models exhibit limited hierarchical feature representation and weak cross-scale interaction due to flat sequential modeling and insufficient spatial inductive biases, leading to sub-optimal performance in complex scenes. To address these issues, we propose a Mamba with Deformable Dilated Convolutions Network (MDDCNet) for accurate traffic object detection in this study. In MDDCNet, a well-designed hybrid backbone with successive Multi-Scale Deformable Dilated Convolution (MSDDC) blocks and Mamba blocks enables hierarchical feature representation from local details to global semantics. Meanwhile, a Channel-Enhanced Feed-Forward Network (CE-FFN) is further devised to overcome the limited channel interaction capability of conventional feed-forward networks, whilst a Mamba-based Attention-Aggregating Feature Pyramid Network (A^2FPN) is constructed to achieve enhanced multi-scale feature fusion and interaction. Extensive experimental results on public benchmark and real-world datasets demonstrate the superiority of our method over various advanced detectors. The code is available at https://github.com/Bettermea/MDDCNet.
Figures & tables
Figure 1: The model architecture of our proposed MDDCNet which comprises Hybrid Backbone (a), Attention-Aggregating Feature Pyramid Network ( A2 FPN) (b) and Detection Head (c).
Figure 2: Illustration of the hybrid backbone network consisting of successive MSDDC and Mamba blocks.
Model variants
MDDCNet-N
MDDCNet-T
MDDCNet-B
Embed. Dim.
[16, 32, 64, 128]
[32, 64, 128, 256]
[64, 128, 256, 512]
Stage 1 (H/4 × W/4)
MSDDC × 3
MSDDC × 3
MSDDC × 3
Stage 2 (H/8 × W/8)
MSDDC × 3
MSDDC × 3
MSDDC × 3
Stage 3 (H/16 × W/16)
Mamba × 9
Mamba × 9
Mamba × 12
Stage 4 (H/32 × W/32)
Mamba × 3
Mamba × 3
Mamba × 3
Params (M)
4.8
6.6
18.0
Table 1: Configurations of different model variants in our MDDCNet family.
Figure 3: Comparison of our proposed CE-FFN (c) and the vanilla FFN (a) and FFN with CA-Block (b).
Figure 4: Our proposed Contextual-Spatial-Channel Attention (CSCA) synergy module. Comprising three complementary attention branches, it leverages an attention-aggregating mechanism for comprehensively strengthening feature discriminative power, which significantly benefits multi-scale feature fusion and interaction.
Figure 5: Statistics of KITTI dataset in traffic object categories (a) and scale distributions (b).
Figure 6: Statistics of our RTOD dataset in traffic object categories (a) and scale distributions (b).
Figure 7: Example images in RTOD dataset.
Detectors
P(%)
R(%)
mAP@50(%)
mAP@50-95(%)
Params(M)
FLOPs(G)
FPS
YOLOv3t
87.7
75.3
82.1
58.3
12.1
18.9
67.52
YOLOv5n
91.3
77.8
88.5
63.3
2.5
7.2
80.64
YOLOv6s
89.8
76.4
85.3
60.6
4.2
11.9
80.65
YOLOv8n
90.5
80.9
89.5
65.1
3.0
8.2
85.32
YOLOv8s
93.1
87.6
93.3
73.0
11.1
28.5
66.67
YOLOv10n
83.7
79.1
85.9
62.5
2.7
8.4
90.50
Table 2: Comprehensive comparison on the KITTI Dataset. The results of our method are highlighted in bold.
Detectors
Car
Van
Truck
Tram
Pedestrian
Cyclist
Misc
YOLOv3t
88.7
86.9
97.3
87.8
68.8
65.8
79.7
YOLOv5n
95.2
91.5
96.9
92.2
76.5
80.7
86.4
YOLOv6s
94.9
87.9
95.6
91.0
76.7
73.9
77.1
YOLOv8n
95.7
93.1
95.8
92.0
79.3
78.6
92.0
YOLOv8s
97.3
96.8
98.8
94.3
84.8
85.8
95.2
YOLOv10n
94.7
90.3
96.2
91.0
73.0
77.0
79.2
Table 3: Comparison of different detectors across various categories on the KITTI dataset (mAP@50%).
Figure 8: PR curves of YOLOv8n (a), YOLOv10n (b), YOLOv11n (c), YOLOv12n (d), YOLOv13n (e) and our MDDCNet-T (f) on the KITTI dataset.
Detectors
P(%)
R(%)
mAP@50(%)
mAP@50-95(%)
FPS
YOLOv5n
87.5
76.1
82.1
57.6
46.18
YOLOv6s
84.3
73.3
79.2
55.2
48.82
YOLOv8n
89.9
79.1
84.3
60.6
55.56
YOLOv9t
83.9
77.3
81.2
57.9
51.28
YOLOv11n
87.1
78.9
84.3
60.1
55.25
YOLOv12n
89.2
77.4
81.7
58.0
52.36
Table 4: Comprehensive comparison of different models on the RTOD Dataset.
Figure 9: PR curves of YOLOv8n (a), YOLOv13n (b), MambaYOLO-T (c), and our MDDCNet-T (d) on the RTOD dataset.
Detectors
Car
Van
Bus
Truck
Person
Cycle
Plate
Traffic light
Traffic sign
Others
YOLOv5n
95.7
87.2
96.9
84.2
71.0
89.1
10.0
98.3
94.5
94.1
YOLOv6s
94.9
79.9
97.8
78.5
67.1
89.8
2.1
97.7
94.1
89.8
YOLOv8n
95.9
87.1
99.4
87.6
72.7
87.5
21.4
97.7
94.1
99.5
YOLOv9t
95.6
87.1
98.6
82.6
67.4
89.9
8.6
97.9
88.3
95.9
YOLOv11n
95.6
86.2
96.8
88.7
71.9
91.1
23.3
97.9
93.5
97.4
YOLOv12n
95.4
86.9
97.7
86.0
70.8
90.0
0.3
98.0
94.9
96.7
Table 5: Comparison of different detectors across various categories on the RTOD dataset (mAP@50%).
Backbone architectures
mAP@50(%)
mAP@50-95(%)
Params(M)
FLOPs(G)
[MSDDC,MSDDC,MSDDC,MSDDC]
91.6
68.4
7.4
13.4
[Mamba,Mamba,Mamba,Mamba]
91.5
69.7
4.9
11.1
[Mamba,Mamba,MSDDC,MSDDC]
91.7
69.1
7.3
12.9
[MSDDC,MSDDC,Mamba,Mamba]
92.1
69.7
5.0
11.6
Table 6: Ablation experiments on backbone structures.
Setting
P(%)
R(%)
mAP@50(%)
mAP@50-95(%)
Params(M)
FLOPs(G)
DR=1,2,3
92.2
84.7
91.6
69.9
5.0
11.6
DR=1,3,5
91.5
84.8
91.9
68.6
5.0
11.6
DR=1,2,4
93.4
83.9
92.1
69.7
5.0
11.6
Table 7: Ablation experiments on the Dilation Rates (DR) in our MSDDC module.
Attention
P(%)
R(%)
mAP@50(%)
mAP@50-95(%)
Params(M)
FLOPs(G)
Concat
92.2
84.7
91.5
69.7
4.9
11.1
MLCA-only
91.6
83.4
91.2
69.4
5.0
11.3
RCSSC
92.8
86.4
91.8
71.3
5.3
12.0
CSCA
93.9
85.4
92.5
71.1
5.5
11.6
Table 8: Ablation experiments on different attention mechanisms.
Methods
P(%)
R(%)
mAP@50(%)
mAP@50-95(%)
Params(M)
FLOPs(G)
Vanilla FFN
92.2
84.7
91.5
69.7
4.9
11.1
CA Block
92.0
84.2
91.7
69.9
5.3
11.5
Residual CA Block
91.0
85.7
91.8
69.4
5.3
11.6
Gated CA Block
91.2
85.3
92.0
69.6
5.5
11.7
CE-FFN
94.2
84.0
92.3
70.0
5.7
11.8
Table 9: Ablation experiments on different FFN structures.
Modules
P(%)
R(%)
mAP@50(%)
mAP@50-95(%)
Params(M)
FLOPs(G)
CE-FFN
MSDDC
CSCA
92.5
84.3
91.5
69.7
4.9
11.1
✓
94.2
84.0
92.3
70.0
5.7
11.8
✓
93.4
83.9
92.1
69.7
5.0
11.6
✓
93.9
85.4
92.5
71.1
5.5
11.6
✓
✓
92.2
85.8
92.8
69.3
6.0
12.2
Table 10: Comprehensive ablation experiments of our MDDCNet model.
Figure 10: Qualitative comparison of YOLOv13n and our MDDCNet on the KITTI dataset. Our MDDCNet demonstrates superior performance with fewer missed detections.
Figure 11: Qualitative comparison of YOLOv13n and our MDDCNet on the RTOD dataset. It can be observed that our method outperforms YOLOv13n by achieving more accurate results with fewer missed targets and lower false positive rates.
Traffic sign detection is a challenging visual signal processing task for advanced driver assistance, where small objects, scale variation, and occlusion limit conventional detectors with fixed receptive fields. This paper proposes State-space Modeling and Dynamic Dual Fusion Network (SMDDFNet), a deep learning detector for traffic sign images. SMDDFNet integrates a Dynamic Dual Fusion (DDF) module and a state-space modeling backbone to enhance multi-scale feature representation. DDF combines efficient multi-scale attention with content-aware dynamic filtering in the frequency domain, while the backbone captures long-range dependencies with linear computational complexity. A multi-scale feature fusion neck further aggregates pyramid features for robust localization of small signs. Experiments on TT100K, GTSDB, PASCAL VOC, and the Roboflow~100 vehicle subset show that SMDDFNet achieves competitive accuracy against recent detectors while retaining real-time throughput. The source code is available at https://github.com/rainbowyuyu/SMDDFNet
TianYi Yu, DaJian Zhong, Lilin Wang
Collage of Infomation Engineering, Shanghai Maritime University, Shanghai, 201306, China
Domain Adaptive Object Detection (DAOD) aims to transfer detectors from a labeled source domain to an unlabeled target domain. Existing DAOD methods employ multi-granularity feature alignment to learn domain-invariant representations. However, the local connectivity of their CNN-based backbone and detection head restricts alignment to local regions, failing to extract global domain-invariant features. Although transformer-based DAOD methods capture global dependencies via attention mechanisms, their quadratic computational cost hinders practical deployment. To solve this, we propose DA-Mamba, a hybrid CNN-State Space Models (SSMs) architecture that combines the efficiency of CNNs with the linear-time long-range modeling capability of State Space Models (SSMs) to capture both global and local domain-invariant features. Specifically, we introduce two novel modules: Image-Aware SSM (IA-SSM) and Object-Aware SSM (OA-SSM). IA-SSM is integrated into the backbone to enhance global domain awareness, enabling image-level global and local alignment. OA-SSM is inserted into the detection head to model spatial and semantic dependencies among objects, enhancing instance-level alignment. Comprehensive experiments demonstrate that the proposed method can efficiently improve the cross-domain performance of the object detector.
Haochen Li, Rui Zhang, Hantao Yao +5
Intelligent Software Research Center, Institute of Software, CAS · University of Chinese Academy of Sciences · State Key Lab of Processors, Institute of Computing Technology, CAS +1
Panoptic segmentation requires the simultaneous recognition of countable thing instances and amorphous stuff regions, placing joint demands on long-range context modelling, multi-scale feature representation, and efficient dense prediction. Existing convolutional and transformer-based methods struggle to satisfy all three requirements concurrently: convolutional architectures are limited in their capacity to model long-range dependencies, while transformer-based methods incur quadratic computational cost that is prohibitive at high resolutions. In this paper, we propose MambaPanoptic, a fully Mamba-based panoptic segmentation framework that addresses these limitations through two principal contributions. First, we introduce MambaFPN, a top-down feature pyramid that leverages Mamba blocks to generate globally coherent, multi-scale feature representations with linear computational complexity. Second, we adopt a PanopticFCN-style kernel generator that produces unified thing and stuff kernels for proposal-free panoptic prediction, enhanced by a QuadMamba-based feature refinement module applied at multiple network stages. Experiments on the Cityscapes and COCO panoptic segmentation benchmarks demonstrate that MambaPanoptic consistently outperforms PanopticDeepLab and PanopticFCN under comparable model sizes, and matches or surpasses Mask2Former on Cityscapes in PQ and AP while requiring fewer parameters.
Qing Cheng, Damiano Bertolini, Wei Zhang +3
Technical University of Munich · Munich Center for Machine Learning (MCML) · Polytechnic University of Milan +3