Out-of-Distribution Segmentation via Wasserstein-Based Evidential Uncertainty

2025-12-12

AI Summary

Open-world semantic segmentation confronts the dual challenges of unknown-class (out-of-distribution, OOD) detection and pixel-wise segmentation. Existing entropy- or confidence-threshold-based approaches exhibit limited discriminative capability for OOD identification. This paper introduces the first integration of the Wasserstein distance into an evidential deep learning framework, leveraging the geometric structure of the probability simplex to model OOD uncertainty. We propose a novel loss comprising Wasserstein distance regularization, Dempster–Shafer evidence fusion, KL divergence–based epistemic regularization, and Dice-based structural consistency constraints. Our method preserves in-distribution (ID) segmentation accuracy while significantly improving OOD detection: it achieves an AUROC gain of over 8% compared to state-of-the-art uncertainty quantification methods and sets new records across multiple benchmarks. The core contribution is a principled, geometrically interpretable, and statistically robust uncertainty modeling paradigm for open-world segmentation.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Segmentation
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

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.

Problem

• Segments unknown objects in open-world scenarios
• Improves safety for automated driving applications
• Uses Wasserstein loss for better OOD segmentation performance

Innovation

• Wasserstein loss for distributional distance capture
• Kullback-Leibler regularization for uncertainty modeling
• Dice structural consistency for segmentation performance