HDMamba-YOLO: Efficient State-Space Perception and Local Spatial Reconstruction for UAV Small Object
Authors: Linduo Wei, Junjie Fan, Yijun Mai, Yong Qi
Organizations: School of Economics and Management, Nanjing University of Science and Technology, Nanjing, 210094, China · School of Intellectual Property, Nanjing University of Science and Technology, Nanjing, 210094, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, 210094, China
Small-object detection in UAV imagery is challenged by weak visual evidence, ambiguous boundaries, dense object distributions, and complex backgrounds. Effective detection therefore requires long-range contextual information for target-background discrimination while preserving explicit local two-dimensional structures for accurate localization. These requirements arise at different stages of the detection pipeline and are not naturally addressed by a uniform feature-processing strategy. We propose Hybrid Dual-domain Mamba-YOLO (HDMamba-YOLO), a stage-wise heterogeneous SSM-CNN detector organized according to a perception-reconstruction-alignment-interaction rationale. EfficientVMamba-based EVSS establishes long-range contextual perception in the backbone, while PhasePatchMerging2D provides phase-aware hierarchical transitions. DST-Wrapper and Native C3k2-ASSAF then perform perception-to-reconstruction transition and repeated local two-dimensional reconstruction during FPN/PAN aggregation. DySample provides content-adaptive cross-scale resampling, while OS-CVTIA introduces macro-micro interaction and task-specific modulation for localization and classification. On VisDrone2019, HDMamba-YOLO-B achieves 42.737% mAP50 and 25.713% mAP50:95 with 10.042M parameters and 29.879 corrected GFLOPs. HDMamba-YOLO-Lite achieves 41.140% mAP50 and 24.741% mAP50:95 with 5.344M parameters. Under the unified AI-TOD evaluation protocol, HDMamba-YOLO-B obtains 21.621% AP and 47.881% AP50. Controlled ablations further support the stage-wise allocation of state-space perception, convolutional reconstruction, dynamic alignment, and task interaction for UAV small-object detection.
Figures & tables
Figure 1: Overall architecture and functional pipeline of HDMamba-YOLO-B. The detector follows a perception–reconstruction–alignment–interaction design: EVSS performs efficient contextual perception and PhasePatchMerging2D constructs phase-aware hierarchical transitions; DST-Wrapper and Native C3k2-ASSAF form an ASSAF-centered local reconstruction pathway; DySample provides content-adaptive cross-scale alignment; and OS-CVTIA performs classification–localization interaction before the native YOLO prediction branches.
Figure 2: P–R–A–I functional rationale and stage-wise operator allocation in HDMamba-YOLO-B. State-space modeling is assigned primarily to long-range contextual perception, CNN-style processing is retained for local two-dimensional reconstruction, DySample provides content-adaptive cross-scale resampling, and OS-CVTIA performs macro–micro interaction for localization and classification. DST-Wrapper acts as the transition between perception and reconstruction. The reconstructed and aligned multi-scale features are P3 ( 80×80×64 ), P4 ( 40×40×128 ), and P5 ( 20×20×192 ). This figure is a functional abstraction rather than a complete layer-wise topology.
Figure 3: Detection-adapted EVSS block used in the contextual perception stage. The block combines ES2D-based selective state-space propagation with a lightweight local path composed of depthwise convolution, Batch Normalization, GELU, and pointwise projection, followed by shared BiAttn calibration, residual updating, and an MLP transformation.
Figure 4: Mechanism of PhasePatchMerging2D. The input feature is decomposed into four parity-defined spatial phases, concatenated in the implementation order [X00,X10,X01,X11] , converted to channel-last layout for LayerNorm and bias-free linear projection, and then returned to channel-first layout. Optional zero padding is applied only to the bottom and/or right boundary when an input dimension is odd.
Figure 5: ASSAF-centered local spatial reconstruction pathway. (a) ASSAF combines directionally factorized local-detail modeling and fixed-grid large-kernel contextual modeling with independent dual-sigmoid calibration and gated additive reconstruction. (b) DST-Wrapper reuses the ASSAF mechanism to re-encode the deepest backbone feature at the perception–reconstruction interface. (c) Native C3k2-ASSAF embeds ASSAF into the native C2f/C3k2 split–iterate–concatenate topology and retains the initial partitions and intermediate ASSAF responses for progressive feature reuse; the canonical HDMamba-YOLO-B configuration uses N=1 .
Figure 6: Mechanism of DySample-based dynamic spatial alignment. A 1×1 offset predictor generates content-dependent residual offsets, which are combined with fixed sub-pixel initialization and the pixel-center base grid. After coordinate normalization and rearrangement, the resulting dynamic grid resamples grouped feature values using bilinear grid_sample . In HDMamba-YOLO-B, g=4 and s=2 ; padding_mode=border and align_corners=False are used.
Figure 7: Structure of OS-CVTIA and its native YOLO prediction interface. Four macro/micro boundary–context responses are blended using α and converted into asymmetric localization and classification masks. The masks modulate the shared feature through learnable residual amplitudes before the native regression ( cv2 +DFL) and classification ( cv3 ) branches. In the final configuration, α=0.5 , λ=0.1 , λα=0.05 , and γbox=γcls=0.01 at initialization.
Dataset
Categories
Training images
Validation images
VisDrone2019
10
6,471
548
AI-TOD
8
11,214
2,804
Table 1: Dataset splits used in the experiments.
Figure 8: Accuracy–complexity comparison on the VisDrone2019 validation set. The horizontal axis denotes reported or corrected GFLOPs, the vertical axis denotes mAP 50:95 , and marker area is proportional to the number of parameters. Marker shapes distinguish detector families, while the connected diamond markers indicate the Tiny, Lite, and Base configurations of HDMamba-YOLO.
Figure 9: Feature-response comparison on representative VisDrone2019 validation scenes. Rows (a)–(c) represent dense mixed traffic and pedestrian clusters, low-contrast vehicles against cluttered road backgrounds, and densely packed distant targets with partial occlusion, respectively. Columns show the input image, YOLO11s response, and HDMamba-YOLO-B response, respectively. Responses are extracted from the P3 pre-detection feature (stride 8) and summarized by channel-wise root-mean-square activation. Within each row, the two models use the same normalization scale; warmer colors indicate larger normalized responses.
Figure 10: Qualitative detection comparison on representative VisDrone2019 validation scenes. The three rows follow the same scene order and challenge categories as Fig. 9 , enabling a direct comparison between feature-response patterns and the corresponding detection outcomes. Columns show the ground-truth annotations, YOLO11s predictions, and HDMamba-YOLO-B predictions, respectively. Orange boxes indicate ground-truth instances. In the prediction panels, green and blue solid boxes denote true-positive (TP) and false-positive (FP) detections, respectively, while red dashed boxes indicate unmatched ground-truth objects (FN). Visual matching uses class-consistent one-to-one assignment at an IoU threshold of 0.5; detections are shown at a confidence threshold of 0.25 with an NMS IoU threshold of 0.7.
Figure 11: Scale-aware qualitative detection comparison on the official AI-TOD validation set. Rows (a)–(c) correspond to a very-tiny-dominated scene (image 1697), a tiny-object scene (image 1043), and a mixed tiny/small scene (image 1845), respectively. Columns show Ground Truth, YOLO11s, and HDMamba-YOLO-B. Orange boxes denote ground-truth instances; green and blue solid boxes in prediction panels denote TP and FP detections, while red dashed boxes denote unmatched ground-truth objects (FN). Purple rectangles indicate the ground-truth-driven regions shown in the enlarged views, and the same region is used across the three columns of each row.
Figure 12: Internal response patterns of the full ASSAF block on two representative VisDrone regions. Columns show the input region, Local-detail response, Context response, and fused ASSAF response. The three response maps are extracted from the same forward pass. For visualization, each response is independently normalized using its 1st–99th percentile range; therefore, color intensity reflects the within-branch spatial distribution rather than absolute activation magnitude across branches.
Detection head
Interaction design
P (%)
R (%)
mAP 50 (%)
mAP 50:95 (%)
Native Detect
–
53.747
40.525
42.388
25.569
OS-CVTIA-H2
Weakened micro interaction
52.032
41.706
42.603
25.533
OS-CVTIA-H1
Full interaction + relaxation
53.367
41.572
42.737
25.713
Table 9: Ablation of OS-CVTIA interaction designs on VisDrone2019.
Figure 13: Training convergence on VisDrone2019 for YOLO11s and HDMamba-YOLO-B. The two panels report validation mAP 50:95 and validation box loss, respectively. Both models exhibit stable optimization, while HDMamba-YOLO-B converges to a higher validation accuracy and a lower box-loss level. The figure is intended to compare final optimization behavior rather than convergence speed.
The rapid development of unmanned aerial vehicle (UAV) technology has made aerial-image object detection increasingly important for natural-resource monitoring, traffic management, and disaster response. Detecting small objects in aerial images remains difficult because objects occupy very few pixels, high-frequency cues are easily lost, and global context is hard to model in cluttered scenes. Existing detectors often retain insufficient edge, corner, and texture information. We propose \ours, an aerial-image detector built on the YOLO12 architecture. The model combines a triple-path high-frequency enhancement convolution module (TriPathHFConv), receptive-field coordinate-attention convolution (RFCAConv), and a Mamba-based global-context module. On VisDrone, at an input resolution of 960*960, ours achieves 60.0% mAP@50 and 38.6%mAP@50:95, demonstrating a favorable accuracy--efficiency trade-off for small-object detection. The benchmark and dataset protocol follow the VisDrone challenge setup.
Small object detection in unmanned aerial vehicle (UAV) imagery is challenging because high-altitude viewpoints produce severe scale variation, weak structural cues, and tight computational budgets. Existing lightweight detectors usually fuse multi-scale features after downsampling, where boundary and texture details have already been attenuated and heterogeneous feature streams may be spatially misaligned. To address these issues, we propose CollabOD, a collaborative detection framework that preserves structural details, aligns cross-path features before fusion, and keeps the detection head lightweight at inference time. CollabOD combines a Dual-Path Fusion Stem, a Dense Aggregation Block, a Bilateral Reweighting Module, and a Unified Detail-Aware Head to strengthen localization-oriented representation while limiting extra computation. On VisDrone, CollabOD obtains 52.4 AP50, 30.8 AP75, and 29.9 AP50:95 with 65.5 GFLOPs; on UAVDT it reaches 31.2 AP50 and 17.4 AP50:95; and on AI-TOD it reaches 45.4 AP50 and 20.0 AP50:95 at 137 FPS. The code is available at: https://github.com/Bai-Xuecheng/CollabOD.
Xuecheng Bai, Yuxiang Wang, Chuanzhi Xu +3
Aviation Traffic Control Technology (SHENZHEN) Co., Ltd., Shenzhen, China · The University of Sydney, NSW, Australia · Research Institute of Traffic Control Technology Co., Ltd., Beijing, China +1
Small object detection in unmanned aerial vehicle (UAV) and remote sensing imagery requires preserving high-resolution detail while modeling long-range context. Adding a stride-4 detection level and removing the stride-32 stage benefits tiny targets but weakens peripheral spatial support, whereas directly inserting selective scanning into the main feature path can interfere with weak local cues. We propose ScopeMamba-YOLO, built around an off-path, zero-gated selective-scanning principle that decouples contextual modeling from the convolutional stream. The principle is instantiated by a Cascaded Global-Context Module (CGCM) in the backbone and a Selective-Scan PAN (SS-PAN) in the neck. An Adaptive Multi-scale Strip (AMS) Block reduces the cost of high-resolution feature extraction, while a Scale-Adaptive DFL (SA-DFL) head reallocates distributional support and regression capacity across scales with only 0.008M additional parameters. Controlled experiments show that matched main-path selective scanning reduces mAP50 by 0.98 pp, whereas off-path CGCM improves the final configuration by 0.67 pp over the three-seed no-CGCM mean; operator controls indicate that this gain is not explained by auxiliary branch capacity alone. ERF analysis further shows that the complete context pathway increases the peripheral energy ratio from 0.008 to 0.090 at stride 8. On VisDrone-2019, ScopeMamba-S achieves 50.8% mAP50 with 3.57M parameters, exceeding YOLOv8s by 10.8 pp while using 32% of its parameters; ScopeMamba-M reaches 52.6% mAP50 with 6.48M parameters. Consistent improvements are also observed on AI-TOD, especially for very-tiny and tiny objects.
Junjie Fan, Yijun Mai, Linduo Wei +5
School of Intellectual Property, Nanjing University of Science and Technology, Nanjing 210094, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China · School of Economics and Management, Nanjing University of Science and Technology, Nanjing 210094, China