A Little Fog for a Large Turn
Generating realistic foggy conditions to test autonomous navigation systems
This project explores how adverse weather conditions, particularly fog, can act as natural adversaries to autonomous navigation systems. We present a novel approach to generating realistic foggy conditions that can be used to test and evaluate autonomous vehicle steering models.
Key Features
- Novel methodology for generating adversarial weather conditions using generative models
- CycleGAN and DistanceGAN-based implementations
- Framework for testing autonomous navigation models under adverse conditions
- Perceptual similarity-based definition of adversarial perturbations
Overview of our fog generation framework showing how clear weather images are transformed into realistic foggy conditions.
Technical Implementation
The project uses two main approaches for generating foggy conditions:
-
CycleGAN Implementation
- Learns mappings between clear and foggy domains
- Maintains structural consistency of the original image
- Generates realistic fog effects
-
DistanceGAN Implementation
- Alternative approach focusing on distance preservation
- Provides different perspective on fog generation
- Requires Python 2 environment
Left: Examples of generated foggy conditions from clear weather images. Right: Analysis of steering angle predictions under foggy conditions.
Impact and Applications
This work has important implications for:
- Testing robustness of autonomous driving systems
- Creating synthetic adverse weather datasets
- Understanding how weather conditions affect model predictions
- Developing more robust navigation systems
The code is available on GitHub and the work was published at WACV 2020. For more details, check out our paper (Machiraju & Balasubramanian, 2020).
