Research map / Detection and segmentation / Open-set and few-shot detection

A Dataset for Semantic Segmentation in the Presence of Unknowns

CVPR 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Interactive and referring segmentation · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ISSU, a real-world Indian-road driving segmentation dataset with 19 known classes plus an anomaly label, static and temporal test splits, and benchmarks state-of-the-art anomaly segmentation methods under closed-set and open-set protocols.

Problem

Existing anomaly segmentation datasets for driving evaluate only anomalies (binary known vs. anomaly) and lack in-domain training data and closed-set plus anomaly labels, so models cannot be evaluated for both knowns and unknowns in realistic, cluttered conditions.

Approach

A fully annotated semantic segmentation dataset from cluttered Indian roads that labels both 19 Cityscapes-style known classes and an anomaly class in test sets, with static and temporal (cross-sensor dashcam video) splits and controlled cross-domain, lighting, and anomaly-size evaluations enabling K+1 open-set evaluation.

Key results

• Mask-level methods trained with OOD data (e.g., RbA, EAM, UNO) reach roughly 95% TPR and FPR under 5% on in-domain static road anomaly evaluation, but performance drops sharply in cross-domain and temporal setups.
• Closed-set IoU on ISSU is about 10% lower than on Cityscapes for in-domain static setup and 20-30% lower for in-domain temporal and cross-domain setups.
• In cross-domain static and temporal setups, UNO detects 79% and 70% of anomalies, but reaching 95% TPR results in FPR above 90%.

Measured effects reported in the paper

Surprising findings

• Pixel-level methods generalize better than mask-level methods from Cityscapes to ISSU in road obstacle evaluation, yielding lower FPR.
• Mask2Former-based methods that perform well in-domain struggle with the temporal sensor shift, showing high FPR.

Limitations

• Anomaly labels are limited to objects on or within 2 m of the road, so unknowns outside the region of interest are not annotated.
• Temporal test set is collected with a consumer dashcam of lower quality, confounding sensor and domain shift.
• Dataset is collected only on Indian roads, limiting geographic diversity.

Experimental setup

Datasets: ISSU-Train Cityscapes

Benchmarks: ISSU-Test-Static ISSU-Test-Temporal

Metrics: AP FPR at 95% TPR TPR at 5% FPR IoU oIoU F1

Baselines: JSR-Net DaCUP PixOOD RbA EAM Pebal UNO Mask2Anomaly

Builds on: Mask2Former Cityscapes SMIYC benchmark

Closest papers by meaning

• MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning (NeurIPS 2024 · similarity 0.88)
• Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.88)
• Generalize or Detect? Towards Robust Semantic Segmentation Under Multiple Distribution Shifts (NeurIPS 2024 · similarity 0.86)
• Open-World Semantic Segmentation Including Class Similarity (CVPR 2024 · similarity 0.86)
• AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.86)
• Self-Supervised Likelihood Estimation with Energy Guidance for Anomaly Segmentation in Urban Scenes (AAAI 2024 · similarity 0.85)