Evaluating the Robustness of Anti-UAV Detection under Controlled Fog Degradation: Fog-Aware Training and Clear-Sky Tradeoff

Gur Levy Birkental; Seyed Sahand Mohammadi Ziabari; Ali Mohammed Mansoor Alsahag

cs.CVSep 10, 2026

Evaluating the Robustness of Anti-UAV Detection under Controlled Fog Degradation: Fog-Aware Training and Clear-Sky Tradeoff

Authors: Gur Levy Birkental, Seyed Sahand Mohammadi Ziabari, Ali Mohammed Mansoor Alsahag

Organizations: University of Amsterdam Amsterdam, The Netherlands · Department of Computer Science and Technology, SUNY Empire State University Saratoga Springs, NY, USA

Abstract

Vision-based anti-UAV systems must function in poor visibility, yet most benchmarks use only clear-sky footage, and previous robustness studies treat adverse weather as a simple present/absent condition. As a result, the impact of fog severity on ground-to-air UAV detection remains poorly understood. This work presents the first severity-controlled fog benchmark for this task: synthetic fog at ten severity levels is applied to the RGB modality of the Anti-UAV300 dataset, comparing a clear-trained YOLOv5m baseline to a fog-aware model trained on both clear and foggy images. Detection performance drops sharply and non-linearly: degradation is front-loaded across light-to-moderate fog (beta approximately 0.05-0.10), with a 96% reduction in [email protected]:0.95 from clear to thickest fog, mainly due to lost recall and confidence. On the comparable metric ([email protected]), this collapse exceeds the most extreme rain degradation reported in the closest prior benchmark. Fog-aware training boosts detection across all severities (up to +0.320 [email protected]:0.95) with only a 9.1% drop in clear-sky accuracy, raising the threshold for reliable detection while not preventing collapse under extreme fog. Since reliability is lost within a narrow visibility range, simple clear vs adverse tests underestimate operational risk.

Figures & tables

Figure 1. Bounding box scale distribution and size category breakdown per split for the RGB modality of the Anti-UAV300 dataset.

Figure 2. Spatial distribution of UAV bounding box centers across all splits.

Figure 3. Baseline YOLOv5m detections on one Anti-UAV300 frame across all ten fog levels ( $\beta$ above each panel). Orange: predictions with confidence; green: ground truth. The target is detected only at $\beta=0.05$ – $0.06$ with falling confidence, then lost from $\beta=0.07$ on, though it remains present throughout.

Figure 4. Per-sequence Pearson and Spearman correlation between bounding-box scale and radial distance from image center (318 sequences).

Figure 5. UAV box centers by size category. Large and medium targets cluster more centrally where radial prior attenuates fog the most; small and tiny targets also mostly cluster centrally, with some outliers.

Figure 6. Training and validation curves for the baseline YOLOv5m model trained on clear sky data. Training halted at epoch 223. Best epoch 123, validation mAP@0.5:0.95 $=0.698$ is used for all inference. Validation clips derive from the same source videos as training, all reported performance uses the held-out test set. — Figure 6. Training and validation curves for the baseline YOLOv5m model trained on clear sky data. Training halted at epoch 223. Best epoch 123, validation [email protected]:0.95 $=0.698$ is used for all inference. Validation clips derive from the same source videos as training, all reported performance uses the held-out test set.

Figure 7. Training and validation curves for the fog-aware YOLOv5m model, trained on a $50{:}50$ mix of clear and fog-augmented data ( $i\in\{0,\ldots,9\}$ , sampled uniformly) for up to 300 epochs with early stopping (patience 100); training halted at epoch 160. Validation mAP@0.5:0.95 peaks at the best checkpoint (epoch $59$ , $0.617$ ) before a slight decline that triggered early stopping. — Figure 7. Training and validation curves for the fog-aware YOLOv5m model, trained on a $50{:}50$ mix of clear and fog-augmented data ( $i\in\{0,\ldots,9\}$ , sampled uniformly) for up to 300 epochs with early stopping (patience 100); training halted at epoch 160. Validation [email protected]:0.95 peaks at the best checkpoint (epoch $59$ , $0.617$ ) before a slight decline that triggered early stopping.

Metric	Baseline	Fog-aware
Precision	$0.971\pm 0.003$	$0.946\pm 0.012$
Recall	$0.919\pm 0.003$	$0.896\pm 0.009$
F1	$0.944\pm 0.001$	$0.920\pm 0.009$
[email protected]	$0.956\pm 0.001$	$0.938\pm 0.009$
mAP0.5:0.95	$0.567\pm 0.002$	$0.515\pm 0.006$

Table 1. Clear-sky test performance of the baseline and fog-aware models (mean

\pm

std across three seeds). Mixed training reduces clear-sky [email protected]:0.95 by

0.052

(

9.1\%

Figure 8. Atmospheric light sensitivity at fog levels $i\in\{2,5,9\}$ for $A\in\{0.3,0.5,0.7\}$ (mAP@0.5:0.95; single representative run per model). — Figure 8. Atmospheric light sensitivity at fog levels $i\in\{2,5,9\}$ for $A\in\{0.3,0.5,0.7\}$ ([email protected]:0.95; single representative run per model).

Fog Severity	Baseline	Fog-aware
Clear	$0.567\pm 0.0017$	$0.515\pm 0.0064$
$\beta=0.05$	$0.373\pm 0.0168$	$0.496\pm 0.0006$
$\beta=0.06$	$0.278\pm 0.0204$	$0.482\pm 0.0020$
$\beta=0.07$	$0.188\pm 0.0165$	$0.463\pm 0.0038$
$\beta=0.08$	$0.122\pm 0.0081$	$0.437\pm 0.0049$
$\beta=0.09$	$0.088\pm 0.0090$	$0.408\pm 0.0056$

Table 2. Mean Average Precision ([email protected]:0.95) across three training seeds for all fog severity levels.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9. Methodology visualization showing the five-stage experimental pipeline, three experimental conditions (clear-sky baseline, fog robustness evaluation, and fog-aware mitigation), and key experimental setup parameters. Star denotes novel contributions.

Split	Data
Training	Clear images
Validation	Clear images
Test	Clear images + fog (all severity levels)

Appendix

Table 3. Data configuration for the clear-sky baseline condition.

Condition	$i$	P	R	F1	mAP @0.5	mAP @0.5:0.95
Clear	–	$0.971\pm 0.003$	$0.919\pm 0.003$	$0.944\pm 0.002$	$0.956\pm 0.001$	$0.567\pm 0.002$
Fog	0	$0.866\pm 0.023$	$0.745\pm 0.028$	$0.801\pm 0.018$	$0.819\pm 0.020$	$0.373\pm 0.017$
Fog	1	$0.800\pm 0.031$	$0.580\pm 0.038$	$0.672\pm 0.028$	$0.679\pm 0.032$	$0.278\pm 0.020$
Fog	2	$0.727\pm 0.025$	$0.382\pm 0.028$	$0.501\pm 0.028$	$0.503\pm 0.032$	$0.188\pm 0.017$
Fog	3	$0.339\pm 0.025$	$0.362\pm 0.030$	$0.349\pm 0.011$	$0.336\pm 0.017$	$0.122\pm 0.008$
Fog	4	$0.320\pm 0.042$	$0.187\pm 0.008$	$0.235\pm 0.007$	$0.229\pm 0.023$	$0.088\pm 0.009$

Appendix

Table 4. Baseline YOLOv5m test-set performance across clear conditions and all ten fog severity levels (

\beta=0.01i+0.05

A=0.5

), reported as mean

\pm

standard deviation across three training seeds. The model is trained on clear data only. From clear to the heaviest fog, [email protected]:0.95 falls from

0.567

0.022

(

96.0\%

), while [email protected] falls from

0.956

0.048

(

95.0\%

), showing that the collapse is not an artefact of the stricter evaluation criterion.

Condition	$i$	P	R	F1	mAP @0.5	mAP @0.5:0.95
Clear	–	$0.946\pm 0.012$	$0.896\pm 0.009$	$0.920\pm 0.009$	$0.938\pm 0.009$	$0.515\pm 0.006$
Fog	0	$0.942\pm 0.008$	$0.862\pm 0.006$	$0.901\pm 0.003$	$0.915\pm 0.004$	$0.496\pm 0.001$
Fog	1	$0.935\pm 0.001$	$0.844\pm 0.004$	$0.888\pm 0.002$	$0.900\pm 0.003$	$0.482\pm 0.002$
Fog	2	$0.926\pm 0.008$	$0.816\pm 0.005$	$0.867\pm 0.001$	$0.882\pm 0.003$	$0.463\pm 0.004$
Fog	3	$0.919\pm 0.005$	$0.790\pm 0.003$	$0.850\pm 0.002$	$0.860\pm 0.003$	$0.437\pm 0.005$
Fog	4	$0.900\pm 0.010$	$0.760\pm 0.002$	$0.824\pm 0.003$	$0.829\pm 0.004$	$0.408\pm 0.006$

Appendix

Table 5. Fog-aware YOLOv5m test-set performance across clear conditions and all ten fog severity levels (

A=0.5

), reported as mean

\pm

standard deviation across three training seeds. The model is trained and validated on a 50:50 clear/fog mixture, with

i\in\{0,\ldots,9\}

. At every fog level, the model outperforms the clear-trained baseline; the largest gain is

+0.320

[email protected]:0.95 at

i=4

Figure 10. Detection performance of the baseline and fog-aware YOLOv5m models on the test set across all ten fog severity levels ( $A=0.5$ ). Yellow line represents the uniform-fog control tested at only 1 seed run , trained using baseline weights.

Figure 11. Absolute performance loss ( $\Delta$ mAP@0.5:0.95) relative to each model’s clear-sky baseline across all ten fog severity levels ( $A=0.5$ ). — Figure 11. Absolute performance loss ( $\Delta$ [email protected]:0.95) relative to each model’s clear-sky baseline across all ten fog severity levels ( $A=0.5$ ).

	Baseline			Fog-aware
Severity	Small	Medium	Large	Small	Medium	Large
Clear	0.528	0.525	0.596	0.448	0.488	0.550
$i{=}0$	0.251	0.334	0.452	0.415	0.455	0.548
$i{=}1$	0.195	0.248	0.353	0.397	0.441	0.540
$i{=}2$	0.153	0.165	0.237	0.371	0.418	0.528
$i{=}3$	0.107	0.093	0.156	0.351	0.386	0.509

Appendix

Table 6. Size-stratified detection performance ([email protected]:0.95) on the test set across clear conditions and all ten fog severity levels for the baseline and fog-aware models. Targets are grouped Small/Medium/Large; the 27 tiny instances are merged into Small (Section 3.1 ).

A=0.5

	Recall		[email protected]:0.95
Severity	640	1280	640	1280
Clear	0.947	0.917	0.528	0.517
$i{=}0$	0.824	0.743	0.251	0.208
$i{=}1$	0.715	0.546	0.195	0.140
$i{=}2$	0.624	0.517	0.153	0.093
$i{=}3$	0.374	0.304	0.107	0.060

Appendix

Table 7. Baseline small-target recall and [email protected]:0.95 at

640

1280

inference resolution across clear conditions and all ten fog severity levels (model trained at

640\times 640

A=0.5

). Inference at

1280

is lower than at

640

at every condition.

Figure 12. True-positive confidence distributions across fog severity (conf $\geq 0.25$ , IoU $\geq 0.5$ ; single representative run per model). The baseline’s confident true positives fall to near-zero from $i\approx 4$ (empty panels reflect detection loss, not missing data); the fog-aware model retains a peak near $0.8$ at every level.

Figure 13. False-discovery rate (FP/(TP+FP), i.e. $1-$ precision) across fog severity (conf $\geq 0.25$ , IoU $\geq 0.5$ ; single representative run). Both models rise with fog; the fog-aware model stays lower from $i=2$ on, at the cost of a marginally higher rate under clear conditions.

Figure 14. Mean IoU of surviving true positives across fog severity (conf $\geq 0.25$ , IoU $\geq 0.5$ ; single representative run; error bars $\pm 1$ std). Localization quality holds at $\sim 0.75$ for both models at every level: surviving detections remain well-localized even as recall collapses.

Figure 15. Precision–recall curves for the baseline and fog-aware models across clear and all ten fog severity levels (single representative run per model). Baseline curves contract toward the origin as fog density rises; the fog-aware model retains substantially more area at every level.

UAV Detection

See all 25 papers

A Task-Driven Evaluation of UAV Detection and Tracking under Synthetic Fog
Jul 6, 2026Amir Pouladi, Vesal Ahsani, Haijun Li +2Visual Object Tracking Aerial Object Detection
Center for Aerospace Research (CfAR), University of Victoria, Victoria, BC, Canada · Advanced Control and Intelligent Systems (ACIS) Lab, University of Victoria, Victoria, BC, Canada
A Data Efficiency Study of Synthetic Fog for Object Detection Using the Clear2Fog Pipeline
May 12, 2026Mohamed Ahmed Mohamed, Xiaowei HuangSynthetic Data Pretraining Synthetic-to-Real Domain Adaptation
School of Computer Science and Informatics, University of Liverpool, Liverpool, L69 3DR, UK
Enhancing Visual Perception in Foggy Conditions via Multiclass Fog Density Modeling
Aug 3, 2026Mohamad Mofeed Chaar, Galia WeidlAutonomous Driving Autonomous Driving Perception
Connected urban mobility, Faculty of Engineering, University of Applied Sciences, Aschaffenburg, Germany

Jul 6, 2026J/K move · Enter open · S save

A Task-Driven Evaluation of UAV Detection and Tracking under Synthetic Fog

Amir Pouladi, Vesal Ahsani, Haijun Li, Homayoun Najjaran+1

Center for Aerospace Research (CfAR), University of Victoria, Victoria, BC, Canada · Advanced Control and Intelligent Systems (ACIS) Lab, University of Victoria, Victoria, BC, Canada

Visual Object Tracking Aerial Object Detection

Open paper AskPDF arXiv

Fog severely degrades the visibility of small unmanned aerial vehicles (UAVs) in skydominant, long-range imagery, reducing the reliability of downstream detection and tracking. This paper presents a task-driven evaluation framework that links depth-aware synthetic fog generation, image restoration, object detection, and tracking within a unified pipeline. Given the practical difficulty of collecting and annotating foggy UAV scenes, synthetic fog is generated from real clear-weather outdoor images containing UAV targets using monocular depth estimation and the atmospheric scattering model. Representative restoration methods from classical, convolutional neural network (CNN)-based, and transformer-based families are first compared, after which the selected restoration model is integrated into the downstream perception pipeline. Detection is evaluated under both clean-only and fog-inclusive training regimes using multiple detector variants, while tracking-by-detection is assessed on clean, foggy, and restored video sequences. Beyond image-level restoration metrics, the study evaluates how fog and restoration affect detection robustness and tracking performance. The results show that fog substantially degrades both detection and tracking, primarily through increased missed detections. Fog-inclusive training provides the most consistent improvement in robustness, whereas test-time restoration is most beneficial when the detector has been trained only on clean imagery. These findings show that restoration quality does not necessarily translate into proportional gains in downstream perception and therefore should be evaluated jointly with detection and tracking performance.

Abstract

UAV Detection

Explore similar work

A Task-Driven Evaluation of UAV Detection and Tracking under Synthetic Fog

A Data Efficiency Study of Synthetic Fog for Object Detection Using the Clear2Fog Pipeline

Enhancing Visual Perception in Foggy Conditions via Multiclass Fog Density Modeling