Learning Uncertainty for Safety-Oriented Semantic Segmentation in Autonomous Driving

Victor Besnier, David Picard, Alexandre Briot

In the high-stakes domain of autonomous driving, a neural network’s confidence can be its most dangerous trait. Standard deep learning models are notoriously overconfident; they may misclassify a glare-obscured obstacle with 99% certainty, potentially leading to catastrophic results. To address this, researchers from Valeo and various French institutes propose a novel "Observer" network designed to flag when a primary segmentation system is likely to fail.

The Problem: Brittle Confidence

Deep learning models for semantic segmentation—identifying every pixel in an image as "road," "pedestrian," or "tree"—are excellent but brittle. They struggle with two primary types of uncertainty:

  1. Epistemic Uncertainty: The "I haven't seen this before" problem, where the model lacks knowledge due to gaps in training data.
  2. Aleatoric Uncertainty: The "The data is messy" problem, caused by sensor noise, motion blur, or environmental factors like heavy rain and sun glare.

Existing solutions like MC Dropout (running the model multiple times to check for variance) are computationally expensive, often exceeding the millisecond-latency requirements of real-time driving. Other methods, such as Distillation, frequently fail to capture the nuances of aleatoric uncertainty. This research seeks a method that is both fast and comprehensive.

The Observer Framework

The core innovation is the Observer Network. Rather than modifying the main segmentation network, the authors attach a secondary network that "observes" the main network’s internal activation maps in parallel.

The Observer architecture runs in parallel with the main segmentation network, utilizing its internal features to estimate uncertainty via KL Divergence.

Uncertainty is defined as the dissimilarity (KL Divergence) between the main network’s output ($\hat{y}$) and the observer’s output ($y_o$). The observer is trained to "disagree" with the main network specifically when the main network is likely to be wrong. This dissimilarity serves as a robust proxy for uncertainty, allowing the system to trigger safety fallbacks when it encounters confusing scenes.

Self-Supervised Training with Oracles

To train the observer without manual labels, the authors utilize two "oracles":

  • Epistemic Oracle: During training, the slow MC Dropout method identifies uncertain pixels. The observer is then taught to maximize disagreement with the main network on these specific pixels.
  • Aleatoric Oracle: The researchers artificially add "sun glare" artifacts to images. The observer is trained to recognize that the main network's prediction on a glare-affected image will likely deviate from the ground truth of a clean image.

Experimental Results

The framework was evaluated on the CamVid dataset, with artificial sun glare introduced to simulate safety-critical failures. The researchers used safety-oriented metrics, such as Recall at 95% Precision (R@P=0.95), to measure how many uncertain situations the system could catch while maintaining high reliability.

Qualitative comparison of uncertainty maps under sun glare. The proposed Observer method (d) captures the glare artifact significantly better than MC Dropout (b) or Distillation (c).

Key Findings:

  • Speed: The framework is approximately 20x faster than high-quality MC Dropout, adding minimal overhead because it utilizes the main network's existing features.
  • Handling Glare: In the presence of sun glare, the Observer achieved a Recall of ~68-79%, whereas standard MC Dropout and Softmax methods plummeted to near 0%.
  • Robustness: On standard out-of-distribution detection, the Observer performed nearly as well as the "gold standard" MC Dropout (T=50) but at a fraction of the computational cost.

Conclusion

The Observer approach demonstrates that learning to "disagree" is a powerful strategy for uncertainty estimation. By capturing both epistemic and aleatoric noise in a single forward pass, it provides a viable path toward safer, real-time autonomous systems. Future work may expand this to other environmental challenges like heavy snow or complex urban occlusions to further validate its robustness across diverse driving conditions.

Summary updated

Abstract

In this paper, we show how uncertainty estimation can be leveraged to enable safety critical image segmentation in autonomous driving, by triggering a fallback behavior if a target accuracy cannot be guaranteed. We introduce a new uncertainty measure based on disagreeing predictions as measured by a dissimilarity function. We propose to estimate this dissimilarity by training a deep neural architecture in parallel to the task-specific network. It allows this observer to be dedicated to the uncertainty estimation, and let the task-specific network make predictions. We propose to use self-supervision to train the observer, which implies that our method does not require additional training data. We show experimentally that our proposed approach is much less computationally intensive at inference time than competing methods (e.g., MC Dropout), while delivering better results on safety-oriented evaluation metrics on the CamVid dataset, especially in the case of glare artifacts.
BibTeX
@inproceedings{besnier2021learning,
  author = {Victor Besnier and David Picard and Alexandre Briot},
  title = {{Learning Uncertainty for Safety-Oriented Semantic Segmentation in Autonomous Driving}},
  booktitle = {ICIP},
  year = {2021},
  url = {https://openreview.net/forum?id=BVfFvCbwlOw},
}

Loading related content...

Figure