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

arXiv:2610.00141 · cs.CV · Submitted 2026-09-10 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Evaluating the Robustness of Anti-UAV Detection under Controlled Fog Degradation".

Tom: Vision-based anti-UAV systems must function in poor visibility, yet most benchmarks use only clear-sky footage, and previous work treats fog as simply present or absent.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So we’ve discussed how this paper sets up a controlled test for anti-UAV detection using synthetic fog at ten severity levels to see how robust our systems are compared to just clear-sky testing. Jane, can you elaborate on the main thesis and why this matters for those who care about UAV security?

Jane: The central thesis is that existing benchmarks fail because they only use clear-sky footage, and this research addresses that by introducing a severity-controlled fog benchmark using synthetic fog applied to the Anti-UAV300 dataset. It claims that detection performance drops sharply and non-linearly under this synthetic fog, and it proposes training a model on both clear and foggy images as a way to mitigate this drop while only incurring a small cost to clear-sky accuracy.

Lu: That controlled environment is key because it allows for a systematic evaluation of detectors under adverse visibility conditions, moving beyond the simple presence or absence tests that previous robustness studies used. It’s about creating a structured testbed for real-world uncertainty.

Meng: From an engineering standpoint, the importance lies in quantifying the impact of fog severity precisely; it moves us from anecdotal evidence to hard data on how performance degrades as visibility worsens across those ten levels. That level of detail is what we need to build dependable defense systems.

Lalam: I think the value is that it establishes a new standard for evaluation, showing that models can be trained to handle this uncertainty without losing their primary function in clear conditions, which speaks volumes about the flexibility of modern AI architectures.

Tom: So, essentially, they are arguing that we need to test our systems against these specific fog levels because the performance drop is significant and non-linear, and they provide a way—fog-aware training—to keep that drop manageable at the expense of only a small clear-sky hit. What's next in this discussion?

Jane: They are highlighting that most reliability is lost within a narrow visibility band, which means simple clear versus adverse evaluations seriously underestimate the actual operational risk faced by anti-UAV systems in environments where visibility fluctuates. This finding directly impacts how we assess the safety and efficacy of our current detection technology.

Lu: It points toward a need for more sophisticated modeling that accounts for gradual degradation rather than just binary states, which is a huge concept in AI research when dealing with complex physical phenomena like atmospheric scattering.

Meng: I see the practical implication as needing to move validation beyond static clear images; we have to simulate the visibility spectrum and train models specifically on those transitions to build systems that are truly resilient.

Lalam: It reinforces the idea that true resilience comes from training models on a diverse, challenging distribution of data, not just the easiest examples.

Tom: Right, so we’ve established that this paper provides a new framework for testing robustness under fog and suggests a specific training strategy to manage the performance trade-off between clear conditions and adverse visibility. This sets the stage for what we're going to look at in their conclusions.

Conclusion: Tom: We’ve got to wrap up our discussion on this paper, "Evaluating the Robustness of Anti-UAV Detection under Controlled Fog Degradation: Fog-Aware Training and Clear-Sky Tradeoff." Jane, can you give us a final summary of the title and what its practical implications are for us listening?

Jane: The authors are providing a severity-controlled benchmark by applying synthetic fog at ten levels to the Anti-UAV300 dataset. Their core message is that while fog causes sharp, non-linear performance drops, training on mixed clear and foggy data provides a good balance, showing that reliability can be maintained under adverse conditions without significantly sacrificing performance in clear skies.

Lu: The implications are significant because it shifts the focus from binary testing to a nuanced understanding of how detection systems handle continuous visibility degradation across different severity levels. It demands that we design AI for resilience against gradual environmental changes.

Meng: From my side, the real-world impact is that this paper gives us a specific methodology to validate our defense systems under realistic, quantifiable adverse conditions, which is essential before we commit these tools to operational use where visibility isn't guaranteed to be perfect.

Lalam: This work suggests that improving AI reliability involves training models on a wider range of environmental challenges rather than just optimizing for ideal conditions, which is a fundamental shift in how we approach building dependable intelligence systems.

Tom: It really boils down to this: reliability doesn't fail all at once; it collapses within a specific visibility band, so two-point testing completely misjudges the actual risk. We need to adopt this more nuanced view when evaluating any vision system for security applications.

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

University of Amsterdam · SUNY Empire State University

cs.CV

Submitted: 2026-09-10

Updated: 2026-09-10

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Vision-based anti-UAV systems must function in poor visibility, yet most benchmarks use only clear-sky footage, and previous work treats fog as simply present or absent.

Key concepts

Severity-Controlled Fog Benchmark: The study created a standardized test for anti-UAV detection by applying synthetic fog at ten distinct severity levels. This moves beyond simple clear/adverse tests to systematically evaluate detector robustness across a spectrum of visibility conditions, providing a more realistic assessment of operational risks.
Fog-Aware Model Training: A model was trained using a dataset that included both clear and synthetically degraded (foggy) images. This training method enhances the model's ability to perform well in poor visibility. The trade-off is minor: the clear-sky accuracy slightly decreases, but detection performance under fog improves significantly.
Non-Linear Performance Degradation: Detection accuracy does not drop smoothly; it collapses sharply, especially in light to moderate fog levels. The worst degradation occurs between clear and thickest fog. This sharp drop is mainly due to a loss of recall and confidence rather than localization errors, as the model still accurately locates targets.
Clear-Sky Tradeoff: The study quantifies the cost of improving performance in poor visibility conditions. Training a model for fog robustness results in only a 9.1% reduction in its clear-sky accuracy. This demonstrates that while fog awareness is beneficial, there is a measurable, though small, penalty to baseline performance under clear conditions.

Terminology

Benchmark — 42× in this paper · explained in 2 episodes
Robustness — 37× in this paper · explained in 5 episodes
Accuracy — 10× in this paper · explained in 1 episode
Aggregation — 3× in this paper · explained in 1 episode
Transformer — 3× in this paper · explained in 2 episodes
Benchmarking — 2× in this paper · explained in 3 episodes
Consistency — 2× in this paper · explained in 3 episodes
Core — 2× in this paper · explained in 2 episodes
Data Augmentation — 2× in this paper · explained in 7 episodes
Distillation — 2× in this paper · explained in 2 episodes
F1 Score — 2× in this paper · explained in 3 episodes
Generalizability — 2× in this paper · explained in 6 episodes
Knowledge Distillation — 2× in this paper · explained in 11 episodes
Representation Learning — 2× in this paper · explained in 2 episodes
Alignment — 1× in this paper · explained in 1 episode
Feature Extraction — 1× in this paper · explained in 2 episodes
Fidelity — 1× in this paper · explained in 2 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes
Generalization — 1× in this paper · explained in 13 episodes
Interpretability — 1× in this paper · explained in 15 episodes
Redundancy — 1× in this paper · explained in 2 episodes
Self-Supervised Learning — 1× in this paper · explained in 7 episodes

Summary

Vision-based anti-UAV systems must function in poor visibility, yet most benchmarks use only clear-sky footage, and previous work treats fog as simply present or absent. The gist: Detection performance drops sharply and non-linearly under synthetic fog, with a 96% reduction in mAP@0.5:0.95 from clear to thickest fog, which is mitigated by training a model on both clear and foggy images at a small cost to clear-sky accuracy.

How it works

The study introduces the first severity-controlled fog benchmark for ground-to-air anti-UAV detection by applying synthetic fog at ten severity levels to the RGB modality of the Anti-UAV300 dataset. This allows for a systematic robustness evaluation of anti-UAV detectors under adverse visibility conditions, moving beyond simple clear versus adverse condition tests. The fog is generated using a physically inspired atmospheric scattering model, where the atmospheric scattering coefficient is varied according to the formula: β = 0.01i + 0.05 for i ∈ [0,..., 9], with atmospheric light fixed at A = 0.5. This formulation functions as a heuristic degradation protocol that applies stronger fog attenuation toward the image center and weaker attenuation near the periphery, utilizing a radial distance prior to approximate depth.

Model Comparison and Training

The research compares two YOLOv5m variants: a clear-trained baseline and a fog-aware model trained on both clear and synthetically degraded images. The fog-aware variant is trained on a 50:50 clear/fog-augmented training set with matching validation distribution. This comparison allows the researchers to assess the robustness gains under fog to be assessed alongside any corresponding reduction in clear-sky performance. The results show that while the fog-aware model boosts detection across all severities (up to +0.320 mAP@0.5:0.95), it incurs only a 9.1% drop in clear-sky accuracy, demonstrating that reliability is lost within a narrow visibility range.

Performance Degradation Analysis

Detection performance degrades non-linearly, with the degradation being front-loaded across light-to-moderate fog (β ≈ 0.05 − 0.10). The baseline model experiences a sharp collapse, with a 96% reduction in mAP@0.5:0.95 from clear to thickest fog, mainly due to lost recall and confidence. This is quantified by the metric: ΔmAP = mAPclear − mAPfog (1), which measures robustness loss across all ten severity levels. Failure analysis confirms that the degradation arises from reduced confidence and recall rather than localization error, as the mean IoU for true positives stays near ∼ 0.75 at all fog levels.

Size-Dependent Degradation and Trade-offs

The study examines whether visibility degradation affects targets differently based on size category (small, medium, large). The results indicate that medium targets degrade fastest, followed by large and small targets in both models. Crucially, the analysis shows that "the fog-aware model preserves the large>medium>small ordering at every severity. Furthermore, testing inference resolution at 1280×1280 on small targets revealed that performance is lower than at 640×640, suggesting the issue is not downscaling but rather the model’s representations learned at 640 × 640 training resolution."

Conclusion and Operational Implication

The study establishes a severity-dependent benchmark for ground-to-air anti-UAV detection under synthetic fog. The primary finding is that most reliability is lost within a narrow visibility band, meaning simple, clear vs adverse evaluations underestimate operational risk. Fog at β = 0.14 harms detection more than extreme rain and exceeds the most severe motion blur from prior work. The fog-aware training improves performance across all levels, but the cost is a small reduction in clear-sky mAP@0.5:0.95 (only 9.1%), suggesting that the clear-sky cost of fog-aware training is small relative to the robustness it provides under fog. The central practical implication is that reliability collapses within a narrow visibility band, so two-point testing misjudges where and how quickly detection fails.

The gist: Detection performance drops sharply and non-linearly under synthetic fog, with a 96% reduction in mAP@0.5:0.95 from clear to thickest fog, which is mitigated by training a model on both clear and foggy images at a small cost to clear-sky accuracy.

Key Findings Enumerated:

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed your provided paper, Evaluating the Robustness of Anti-UAV Detection under Controlled Fog Degradation. The core contribution is establishing a severity-controlled benchmark for fog robustness in anti-UAV detection using synthetic fog augmentation and a YOLOv5m architecture.

Based on this research, here are the specific improvements I can propose for AI systems and what those improved systems will be capable of:

The system should be upgraded to incorporate a Fog-Aware Training strategy, specifically training the model on a 50:50 mixture of clear and fog-augmented images.
This improved system will exhibit significantly higher Mean Average Precision (mAP) under adverse visibility conditions compared to a standard clear-sky trained baseline model, with the largest gain occurring at moderate fog severity (+0.320 mAP@0.5:0.95 at severity level 4).
The improved system will demonstrate a stable performance profile across all ten controlled fog severity levels (i=0 to i=9), effectively mitigating the sharp, non-linear performance collapse observed in clear-sky models under fog.
The system can reliably detect UAVs even at heavy fog (severity level 9), achieving a detection rate of 0.093 mAP@0.5:0.95, compared to near-zero performance for the baseline model (0.022).
The improved system will maintain high localization quality (Mean IoU 75) even when its recall drops significantly under fog, indicating that the mitigation strategy successfully suppresses false detections and maintains box precision rather than displacing them.
The system can be deployed with a quantifiable Clear-Sky Cost of only 9.1% reduction in performance during clear conditions, which is acceptable for safety-critical applications where robustness is paramount.
The improved system will allow operators to identify the precise Critical Visibility Threshold—the narrow range of moderate fog (i ≈ 4) where reliability is most likely lost, enabling proactive risk assessment and decision-making based on real-time environmental conditions, rather than relying on a binary clear/fog classification.
The system can be used for size-stratified analysis, allowing operators to understand that large targets remain the most robust through moderate fog (i ≤ 7), while medium targets degrade fastest, guiding tactical prioritization of which threats require immediate attention during visibility reduction.

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 mAP@0.5:0.95 from clear to thickest fog, mainly due to lost recall and confidence. On the comparable metric (mAP@0.5), 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 mAP@0.5:0.95) 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.