Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Soiling detection for Advanced Driver Assistance Systems

Filip Beránek, Václav Diviš, Ivan Gruber

TL;DR

This work addresses robust soiling detection for automotive cameras by framing it as semantic segmentation and comparing multiple segmentation networks to tile-based baselines. It scrutinizes the Woodscape dataset, uncovering data leakage and annotation issues, and contributes a refined sequence-based data split along with four annotation-strict training subsets. Empirical results show semantic segmentation outperforming tile-level methods, with FPNet achieving about 94% accuracy, and reveal limited gains from increasing encoder size due to the dataset's modest scale. The paper emphasizes data-centric improvements and suggests future exploration of hyperparameters and transformer-based approaches to further enhance robustness in adverse conditions.

Abstract

Soiling detection for automotive cameras is a crucial part of advanced driver assistance systems to make them more robust to external conditions like weather, dust, etc. In this paper, we regard the soiling detection as a semantic segmentation problem. We provide a comprehensive comparison of popular segmentation methods and show their superiority in performance while comparing them to tile-level classification approaches. Moreover, we present an extensive analysis of the Woodscape dataset showing that the original dataset contains a data-leakage and imprecise annotations. To address these problems, we create a new data subset, which, despite being much smaller, provides enough information for the segmentation method to reach comparable results in a much shorter time. All our codes and dataset splits are available at https://github.com/filipberanek/woodscape_revision.

Soiling detection for Advanced Driver Assistance Systems

TL;DR

This work addresses robust soiling detection for automotive cameras by framing it as semantic segmentation and comparing multiple segmentation networks to tile-based baselines. It scrutinizes the Woodscape dataset, uncovering data leakage and annotation issues, and contributes a refined sequence-based data split along with four annotation-strict training subsets. Empirical results show semantic segmentation outperforming tile-level methods, with FPNet achieving about 94% accuracy, and reveal limited gains from increasing encoder size due to the dataset's modest scale. The paper emphasizes data-centric improvements and suggests future exploration of hyperparameters and transformer-based approaches to further enhance robustness in adverse conditions.

Abstract

Soiling detection for automotive cameras is a crucial part of advanced driver assistance systems to make them more robust to external conditions like weather, dust, etc. In this paper, we regard the soiling detection as a semantic segmentation problem. We provide a comprehensive comparison of popular segmentation methods and show their superiority in performance while comparing them to tile-level classification approaches. Moreover, we present an extensive analysis of the Woodscape dataset showing that the original dataset contains a data-leakage and imprecise annotations. To address these problems, we create a new data subset, which, despite being much smaller, provides enough information for the segmentation method to reach comparable results in a much shorter time. All our codes and dataset splits are available at https://github.com/filipberanek/woodscape_revision.

Paper Structure

This paper contains 10 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset analysis
  4. Data split
  5. Class definition revision
  6. Annotation revision
  7. Experiments
  8. Networks experiments
  9. Training sets experiments
  10. Conclusion

Figures (4)

  • Figure 1: Example of a soiling annotation from the Woodscape dataset. Black is Clear, green is Transparent, blue is Semi-transparent, and red is Opaque class
  • Figure 2: Example illustrating the issues with annotations. The first row exhibits a clear annotation error, while the second row showcases a simplification in class differentiation.
  • Figure 3: Measured test accuracies across all networks and different training sets.
  • Figure 4: Measured training time per epoch across all networks and different training sets.