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

Christoph Gerhardt, Muhammad Momin Salman, Wolfgang Broll · TU Ilmenau, Germany

Related Repositories

Open-source tools developed as part of the AWARE project.

SWIFT

Weather Classification

PRISM

Image Quality Metrics

PROVE

Downstream Training & Evaluation

Key Findings

Three surprising results from evaluating 21 augmentation strategies on adverse weather segmentation.

21
Augmentation strategies
5
Model architectures
6
Real-world datasets
71.4k
AWACS images
3.1×
Diversity advantage
1

Training Data Diversity Dominates

Training data diversity yields 3.1× larger improvements than any single augmentation strategy. Adding diverse real-world datasets provides +7.08 pp mIoU, while the best augmentation adds +2.27 pp.

2

Quality–Performance Paradox

Methods that produce visually better images (lower FID, higher SSIM) actually yield worse downstream performance. Aggressive transformations that distort appearance improve robustness more than photorealistic synthesis.

3

Augmentation Substitution Effect

Synthetic and real adverse weather data compete rather than complement. When training data already includes real adverse conditions (Stage 2), augmentation gains nearly vanish, falling from +2.27 to +0.65 pp.

AWARE Pipeline Architecture

Interactive overview of the complete AWARE evaluation pipeline — from data curation through generative augmentation to downstream evaluation. Hover over any component for details, click to trace its full data path.

AWACS Dataset

Adverse Weather Augmentation Collection from Segment Anything 1 Billion — 71,400 balanced images across 7 weather conditions.

71.4k
Total images
10.2k
Per condition
7
Weather conditions
SA-1B
Source dataset

Average Color Profiles per Condition

Each stripe shows the average color distribution of images classified into that weather condition, revealing distinct visual signatures.

Clear Day
Clear Day
Cloudy
Cloudy
Dawn/Dusk
Dawn/Dusk
Fog
Fog
Night
Night
Rainy
Rainy
Snowy
Snowy
📦 Download AWACS (coming soon)

Strategy Leaderboard

Stage 1 performance ranking. Click column headers to sort. All values are mIoU (mean Intersection over Union).

Loading leaderboard data...

Citation

If you find this work useful, please cite our paper.

@article{aware2025,
  title={Enhancing Computer Vision Robustness in Adverse Weather 
         Conditions through Advanced Data Augmentation},
  author={[Author Name]},
  journal={IEEE Access},
  year={2025},
  volume={},
  pages={},
  doi={}
}