Small Object Detection in Drone Aerial Imagery with LAF-YOLOv10

Quratulain Nayeem; Fahmina Taranum; Mohammed Mudassir Uddin

cs.CVSep 13, 2026

Small Object Detection in Drone Aerial Imagery with LAF-YOLOv10

Authors: Quratulain Nayeem, Fahmina Taranum, Mohammed Mudassir Uddin

Abstract

General-purpose object detectors lose accuracy on UAV footage, where targets span only a handful of pixels and onboard compute is limited. Prior work composes independently-validated architectural techniques into one detector, assuming gains reported in isolation transfer once combined. We stress-test that assumption directly. LAF-YOLOv10 integrates four techniques into YOLOv10n: a Partial Convolution C2f (PC-C2f) backbone block, an Attention-Guided Feature Pyramid Network (AG-FPN), a P2 detection head replacing the large-object P5 head, and Wise-IoU v3 regression, asking whether their combined effect matches what each contributes alone. We train LAF-YOLOv10 three times (seeds 42, 123, 256) on VisDrone-DET2019, benchmark against unmodified YOLOv10n, and use TIDE error decomposition, per-category breakdown, per-component ablation, attention/loss comparisons, zero-shot transfer to UAVDT, and held-out/test-dev evaluation to localize where the combination succeeds or fails. Composability does not hold here. LAF-YOLOv10 reaches 24.0+/-0.4% [email protected] at 2.14M parameters, 7.8 points below YOLOv10n (31.8%), a deficit that transfers to UAVDT (-10.0 points) and is confirmed by held-out and test-dev evaluation (23.5%, 22.5%). Background false positives, localization error, and duplicate detections move in the direction AG-FPN and Wise-IoU were designed to push. Ablation traces the deficit to a specific source: the P2/-P5 head swap costs 2.5 points independently plus a 2.5-point interaction penalty when layered onto a backbone weakened by PC-C2f, whose own 2.0-point loss is consistent with a partial pretrained-weight transplant (73/150 backbone tensors transfer). The failure is attributable to a specific interaction, not the four components individually. Composability must be verified directly, not assumed. Code/checkpoints: https://github.com/Mudassiruddin7/Small-Object-Detection-in-UAV-Imagery.

Figures & tables

Panel of Figure 1 — Figure 1: LAF-YOLOv10 architecture. PC-C2f backbone (blue), SE-gated AG-FPN neck (purple), and P2/P3/P4 detection heads (green); P5 is removed and Wise-IoU v3 [ 21 ] (Eq. 5-6) supervises regression. Dashed blue arrows denote SE-gated lateral connections; solid purple arrows denote DySample-upsampled top-down fusion.

Panel of Figure 2 — Figure 2: Internal module structure. (a) PC-C2f block [ 48 ] (Eq. 1-2) with $C/4$ active channels. (b) AG-FPN fusion step (Eq. 3-4) via SE gating [ 17 ] and DySample [ 18 ] upsampling. Panels (a) and (b) occupy disjoint horizontal ranges so no label from either can overlap the other.

Method	mAP @0.5	mAP @[.5,.95]	P (%)	R (%)	Params (M)	GFLOPs	FPS
YOLOv5n	27.4	14.8	38.1	28.2	1.9	4.5	156
YOLOv8n	32.1	19.0	42.8	31.6	3.0	8.1	128
YOLOv10n	31.8	18.5	42.6	31.7	2.7	8.2	136
YOLO11n	32.2	18.4	42.9	32.3	2.6	6.3	163
BGF [ 14 ]	32.1	19.5	41.5	32.0	2.0	8.6	37
OSD [ 15 ]	33.4	19.1	43.9	32.5	1.6	7.9	136

Table 1: Results on VisDrone-DET2019 [ 2 ] validation set (640

\times

640). Best in bold , second underlined (LAF-YOLOv10 excluded from ranking on mAP given the result below). Every row except LAF-YOLOv10 is copied from its own original publication rather than retrained here. *LAF-YOLOv10’s P/R use a different protocol (single threshold, greedy match) than the literature values and are not directly comparable; see prose.

Method	[email protected]	mAP@[.5,.95]	Params (M)	GFLOPs
YOLOv10n	47.8	27.3	2.7	8.2
LAF-YOLOv10	37.8	19.8	2.14	13.55
$\Delta$	$-$ 10.0	$-$ 7.5	–	–

Table 2: Zero-shot cross-dataset transfer to UAVDT [ 3 ] (car, bus, truck), models trained on VisDrone-DET2019 only.

Configuration	P	R	mAP	mAP	Params	GFL-
	(%)	(%)	@0.5	@[.5,.95]	(M)	OPs
Baseline	42.3	31.4	31.5	18.2	1.98	6.28
+ PC-C2f only	40.8	29.9	29.5	16.8	1.62	5.41
+ AG-FPN only	43.6	32.4	32.5	19.0	2.00	6.32
+ P2/ $-$ P5 only	44.8	29.6	29.0	16.5	1.65	8.08
+ WIoU only	43.1	31.9	32.0	18.6	1.98	6.28

Table 3: Single-module ablation, each added independently to a from-scratch baseline (seed 42, 300 epochs).

Configuration	P	R	mAP	mAP	Params	GFL-
	(%)	(%)	@0.5	@[.5,.95]	(M)	OPs
Baseline	42.3	31.4	31.5	18.2	1.98	6.28
+ PC-C2f	40.8	29.9	29.5	16.8	1.62	5.41
+ AG-FPN	42.4	30.8	30.5	17.5	1.64	5.46
+ P2 Head ( $-$ P5)	55.2	37.8	25.5	13.5	1.31	7.33
+ WIoU (Full)	56.1	38.4	26.0	13.8	1.31	7.33

Table 4: Additive ablation, components stacked incrementally (seed 42, 300 epochs).

Panel of Figure 3 — Figure 3: Real training convergence for LAF-YOLOv10 on VisDrone-DET2019 [ 2 ] (seed 42, validation [email protected] every 5 epochs, read from the training log).

Category	YOLOv10n	LAF-YOLOv10	$\boldsymbol{\Delta}$
Pedestrian	33.5	31.8	$-$ 1.7
People	24.8	22.5	$-$ 2.3
Bicycle	9.2	6.2	$-$ 3.0
Car	73.0	70.0	$-$ 3.0
Van	39.8	24.8	$-$ 15.0
Truck	28.4	13.6	$-$ 14.8

Table 5: Per-category [email protected] (%) on VisDrone-DET2019 [ 2 ] .

\boldsymbol{\Delta}

is LAF-YOLOv10 minus YOLOv10n; YOLOv10n is a literature value, LAF-YOLOv10 is a real 3-seed mean measured here.

Panel of Figure 4 — Figure 4: Per-category [email protected] on VisDrone-DET2019 [ 2 ] . LAF-YOLOv10 falls below YOLOv10n in every category, most sharply on medium-to-large vehicles.

Method	Cls	Loc	Both	Dupe	Bkg	Miss
YOLOv10n	3.8	5.2	1.4	0.6	8.1	12.3
LAF-YOLOv10	11.5	2.7	0.1	0.05	0.04	27.7
$\Delta$	$+$ 7.7	$-$ 2.5	$-$ 1.3	$-$ 0.55	$-$ 8.06	$+$ 15.4

Table 6: TIDE [ 22 ] error decomposition on VisDrone-DET2019 [ 2 ] ([email protected] penalty, %; lower is better). The YOLOv10n row is a literature value; TIDE is run once on the full model against the validation set, not separately per component.

Split	Images	[email protected]	mAP@[.5,.95]
Validation (reported above)	548	24.0 $\pm$ 0.4	12.3 $\pm$ 0.2
Held-out internal	647	23.5	11.9
Test-dev	1,610	22.5	11.4

Table 7: LAF-YOLOv10 across splits, each evaluated once after all design decisions were frozen.

Aspect	Finding
Headline result	24.0 $\pm$ 0.4% [email protected], $-$ 7.8 pts vs. YOLOv10n
Interaction penalty	$-$ 2.5 pts beyond the sum of independent effects (Table 3 – 4 )
Dominant residual error	Missed detections (TIDE, $+$ 15.4-pt penalty)
Secondary residual error	Classification confusion ( $+$ 7.7-pt penalty)
What improved anyway	Background FPs, localization, duplicates (TIDE, all down)
Worst-hit categories	Van, truck, bus, awning-tricycle ( $-$ 10 to $-$ 15 [email protected])

Table 8: Summary of findings.

Explore similar work

Sep 12, 2026cs.CV

YOLO12-MambaScan: An Efficient Object Detector with High-Frequency Enhancement and State-Space Modeling

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.

Hao Wang

BDNRC

Aerial Object Detection Remote Sensing Small Object Detection Efficient Object Detection

Ask PDF

Sep 19, 2026cs.CV

HDMamba-YOLO: Efficient State-Space Perception and Local Spatial Reconstruction for UAV Small Object

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.

Linduo Wei, Junjie Fan, Yijun Mai +1

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

Aerial Object Detection Remote Sensing Small Object Detection State Space Models

Ask PDF

Mar 6, 2026cs.CV

CollabOD: Collaborative Multi-Backbone with Cross-scale Vision for UAV Small Object Detection

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

Remote Sensing Small Object Detection Small Object Detection Efficient Object Detection Multi-Scale Feature Fusion Real-Time Object Detection Aerial Object Detection

Ask PDF