Abstract
Deep neural networks achieve superior performance in semantic segmentation, but are limited to a predefined set of classes, which leads to failures when they encounter unknown objects in open-world scenarios. Recognizing and segmenting these out-of-distribution (OOD) objects is crucial for safety-critical applications such as automated driving. In this work, we present an evidence segmentation framework using a Wasserstein loss, which captures distributional distances while respecting the probability simplex geometry. Combined with Kullback-Leibler regularization and Dice structural consistency terms, our approach leads to improved OOD segmentation performance compared to uncertainty-based approaches.
Introduction
Deep Neural Networks (DNNs) are excellent at "closed-set" tasks, such as identifying cars, roads, and pedestrians in urban environments. However, they struggle in "open-world" scenarios. When a standard model encounters an object it has never seen before—like a stray animal or a unique road obstacle—it often attempts to force that object into a known category with dangerously high confidence. This limitation poses significant risks for safety-critical applications like autonomous driving.
Existing approaches to mitigate this issue generally fall into two categories: Bayesian/sampling methods (e.g., MC Dropout or Ensembles) and Softmax-based metrics. While sampling methods can be accurate, they are computationally expensive, requiring multiple forward passes. Conversely, Softmax-based metrics are efficient but notoriously overconfident on out-of-distribution (OOD) data. This research aims to develop a method that is both computationally efficient (single forward pass) and mathematically principled to provide reliable uncertainty estimates.
Methodology
The proposed framework builds on Evidential Deep Learning (EDL). Instead of predicting a simple probability for each class, the network predicts "evidence" values that define a Dirichlet distribution. This allows the model to represent not just the predicted class, but also the total uncertainty associated with each pixel.
Wasserstein-Based Loss
The core innovation is the replacement of traditional Euclidean loss functions with a Wasserstein-based evidential loss. Traditional EDL often utilizes Mean Squared Error (MSE), which treats the probability space as a flat Euclidean surface. This approach tends to push predictions toward the "corners" of the probability simplex (the classes), inadvertently causing overconfidence.
By using the Wasserstein metric, the loss function accounts for the underlying geometry of the probability simplex. It centers uncertainty between class boundaries rather than shifting it toward a specific class. To optimize this in practice, the authors employ a "Loss Cocktail" combining:
- Wasserstein Loss: To respect simplex geometry.
- Dice Loss: To ensure spatial coherence, grouping neighboring pixels into consistent labels.
- Annealed KL Regularization: To penalize overconfidence in ambiguous regions.
- MSE Component: A small, temporary component used to stabilize early training stages.
Experiments and Results
The model was trained on the Cityscapes dataset and evaluated using the SegmentMeIfYouCan benchmark, which includes real-world OOD challenges such as LostAndFound (LAF) (small road obstacles) and RoadObstacle21 (diverse OOD objects).
Performance Comparison
The proposed method consistently outperformed standard baselines, including Maximum Softmax Probability and Entropy, as well as complex sampling methods like MC Dropout.
On the RoadAnomaly21 dataset, the Wasserstein-based approach achieved an AuPRC (Area under Precision-Recall Curve) of 54.3, nearly doubling the performance of the standard Entropy baseline (30.0). Furthermore, the model demonstrated high efficiency, requiring only a single forward pass, and showed a low calibration error (0.0347) on in-distribution data, indicating that its confidence levels accurately reflect its predictive accuracy.
Discussion and Key Insights
The research highlights that for OOD detection, the method of measuring distances between distributions is more critical than point accuracy. The Wasserstein distance is "distribution-aware," providing a more nuanced understanding of uncertainty than MSE. Additionally, the inclusion of structural consistency (Dice loss) is vital; since OOD objects are physical entities, grouping uncertain pixels into coherent objects significantly improves segment-level performance.
Limitations and Future Work
While effective, the performance of this framework is sensitive to the balancing of its four loss terms ($\lambda_W, \lambda_{Dice}, \lambda_{KL}, \lambda_{MSE}$). Additionally, the Wasserstein objective can lead to slower convergence compared to standard cross-entropy. Future work may explore extending this framework to class-incremental learning, allowing models to not only detect OOD objects but also learn to incorporate them as new classes over time.
Summary updated