A Little Fog for a Large Turn

Generating Foggy Scenes for Testing Autonomous Navigation Models

WACV 2020

Harshitha Machiraju, Vineeth N Balasubramanian
IIT Hyderabad


Small, carefully crafted perturbations called adversarial perturbations can easily fool neural networks.
sample perturbation

However, these perturbations are largely additive and not naturally found. We turn our attention to the field of Autonomous navigation wherein adverse weather conditions such as fog have a drastic effect on the predictions of these systems. These weather conditions are capable of acting like natural adversaries that can help in testing models.
To this end, we introduce a general notion of adversarial perturbations, which can be created using generative models and provide a methodology inspired by Cycle-Consistent Generative Adversarial Networks to generate adversarial weather conditions for a given image.
Our formulation and results show that these images provide a suitable testbed for steering models used in Autonomous navigation models. Our work also presents a more natural and general definition of Adversarial perturbations based on Perceptual Similarity.

Deviation seen in Popular Steering Models


Fooling Models: Ground truth Steering Angle (in radians) for each of the original test samples. The angles right below indicate the ordered pair of predicted steering angle by AutoPilot and Comma AI respectively. From the second row onward, we indicate the image translation model used and respective steering model it was trained on. The angle below each of those images indicates the prediction by the steering model for the generated foggy image.


paper thumbnail

Arxiv (in WACV'20): https://arxiv.org/abs/2001.05873

Citation: Bibtex

Download Code

You can download our implementation from GitHub.

Visit Github Page


For any queries please contact:

Harshitha Machiraju (ee14btech11011@iith.ac.in)

Vineeth Balasubramanian(vineethnb@iith.ac.in)