EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

AWARE: Augmentation for Weather-Adverse Robustness Evaluation—Benchmarking Strategies for Improved Semantic Segmentation Under Challenging Outdoor Conditions

Christoph Gerhardt, Muhammad Momin Salman, Wolfgang Broll

IEEE Access · 2026 · 14 · 89643-89675

DOI: 10.1109/access.2026.3699854 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Christoph Gerhardt und 2 weiteren Autor:innen. Die Arbeit nennt 100 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Deutsche Forschungsgemeinschaft.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Among recent advances in AI, the progress in synthetic image generation stands out as particularly notable. State-of-the-art models can synthesize controllable photorealistic images that are nearly indistinguishable from real photographs. A natural application of these new capabilities is data augmentation—generating diverse training samples to compensate for scarce or imbalanced real-world data. Yet whether higher-fidelity synthetic images actually translate into better downstream task performance remains unclear. This paper addresses this gap through a large-scale evaluation of 21 generative augmentation strategies—spanning GANs, diffusion models, style transfer, 2D rendering, and multimodal approaches—focused specifically on semantic segmentation as a representative dense prediction task. The image generation is based on AWACS, a new weather-balanced dataset with 71,400 images. To enable principled comparison, we present three main components: SWIFT for condition-aware dataset characterization, PRISM for standardized perceptual and semantic quality assessment, and PROVE for downstream task evaluation. Each strategy is tested across four architectures and six real-world datasets (four for primary evaluation, two for cross-dataset generalization) in a three-stage protocol isolating augmentation effects. Three counter-intuitive findings emerge: 1) adding real weather diversity to training yields modest overall mIoU gains (+2.45 pp) but disproportionately large adverse-condition gains (+6.04 pp), approximately $2.2\times $ larger than the best generative strategy (+2.70 pp adverse), establishing real data diversity—not synthesis—as the dominant lever for adverse-weather robustness, though this comparison involves different effective data budgets; 2) a quality–performance paradox where perceptually worse images drive better segmentation ( $\rho = +0.594$ , $p = 0.005$ ); and 3) an augmentation substitution effect where synthetic and real weather data compete for the same representational capacity, compressing gains by 78–99% once diverse real data is available. Generative augmentation is thus most valuable as a surrogate for missing weather diversity, not as a universal performance booster.

Förderung

  • Deutsche Forschungsgemeinschaft

Zitieren

APA

Gerhardt, C., Salman, M. M., & Broll, W. (2026). AWARE: Augmentation for Weather-Adverse Robustness Evaluation—Benchmarking Strategies for Improved Semantic Segmentation Under Challenging Outdoor Conditions. IEEE Access, 14, 89643-89675. https://doi.org/10.1109/access.2026.3699854

Vancouver

Gerhardt C, Salman MM, Broll W. AWARE: Augmentation for Weather-Adverse Robustness Evaluation—Benchmarking Strategies for Improved Semantic Segmentation Under Challenging Outdoor Conditions. IEEE Access. 2026;14:89643-89675. doi:10.1109/access.2026.3699854

BibTeX
@article{Gerhardt2026,
  author = {Gerhardt, Christoph and Salman, Muhammad Momin and Broll, Wolfgang},
  title = {AWARE: Augmentation for Weather-Adverse Robustness Evaluation—Benchmarking Strategies for Improved Semantic Segmentation Under Challenging Outdoor Conditions},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {89643--89675},
  doi = {10.1109/access.2026.3699854}
}
Kostenlos starten →