SupLID: Geometrical Guidance for Out-of-Distribution Detection in Semantic Segmentation

TL;DR

This work tackles pixel-level OOD detection in semantic segmentation by identifying limitations of traditional classifier-based confidence scores, such as overconfidence and reliance on decision boundaries. It introduces SupLID, a post-hoc framework that guides classifier-derived OOD scores with geometry-derived cues from Local Intrinsic Dimensionality (LID) via a geometrical coreset and superpixel aggregation, enabling scalable, real-time OOD localization without retraining. The approach achieves state-of-the-art performance across multiple benchmarks (e.g., SMIYC, Fishyscapes, RoadAnomaly) and consistently boosts existing retraining-based OOD methods, while preserving the base segmentation quality. Practically, SupLID offers a plug-and-play enhancement for deployment in open-world settings, with code provided for reproducibility and further extensions to distance-based geometrical scoring.

Abstract

Out-of-Distribution (OOD) detection in semantic segmentation aims to localize anomalous regions at the pixel level, advancing beyond traditional image-level OOD techniques to better suit real-world applications such as autonomous driving. Recent literature has successfully explored the adaptation of commonly used image-level OOD methods--primarily based on classifier-derived confidence scores (e.g., energy or entropy)--for this pixel-precise task. However, these methods inherit a set of limitations, including vulnerability to overconfidence. In this work, we introduce SupLID, a novel framework that effectively guides classifier-derived OOD scores by exploiting the geometrical structure of the underlying semantic space, particularly using Linear Intrinsic Dimensionality (LID). While LID effectively characterizes the local structure of high-dimensional data by analyzing distance distributions, its direct application at the pixel level remains challenging. To overcome this, SupLID constructs a geometrical coreset that captures the intrinsic structure of the in-distribution (ID) subspace. It then computes OOD scores at the superpixel level, enabling both efficient real-time inference and improved spatial smoothness. We demonstrate that geometrical cues derived from SupLID serve as a complementary signal to traditional classifier confidence, enhancing the model's ability to detect diverse OOD scenarios. Designed as a post-hoc scoring method, SupLID can be seamlessly integrated with any semantic segmentation classifier at deployment time. Our results demonstrate that SupLID significantly enhances existing classifier-based OOD scores, achieving state-of-the-art performance across key evaluation metrics, including AUR, FPR, and AUP. Code is available at https://github.com/hdnugit/SupLID.

Figures (7)

• Figure 1: Pixel-wise OOD scores overview. Example OOD score maps are shown for a SOTA segmentation model (DeepLabv3+ by NVIDIA DeepLab_NVIDIA) and a re-trained model for OOD detection (PEBAL PEBAL) using a classifier-based score (energy), followed by OOD scores generated using LID based on the proposed coreset. The two methods exhibit complementary properties: LID assigns lower scores to false positive regions and signals missed anomaly regions, but may fail to capture full anomalous objects due to its local nature. Our method (last column) produces more robust and complete OOD maps compared to both approaches.
• Figure 2: Summary of SupLID performance and other OOD scores on five OOD benchmarks across different models.
• Figure 3: Compatibility of SupLID with various classifier-based scores across different baselines on SMIYC-Anomaly dataset.
• Figure 4: OOD prediction maps at pixel and superpixel levels based on energy EnergyOE of the baseline. Superpixel-level aggregation produces sharper anomaly boundaries and improves the consistent classification of neighboring object pixels.
• Figure 5: Variation of OOD detection performance across different values of $k$ for various coreset sizes $m$. Averaged over six random runs on SMIYC-Anomaly.

Theorems & Definitions (1)

• definition 1: Local Intrinsic Dimensionality LID