A Controlled Corruption Dataset for Autonomous Driving (DRIVE-C)

Abstract

DRIVE-C is a controlled corruption dataset for autonomous driving research, designed to support studies of perception robustness, degradation awareness, and sensor health estimation. It uses real-world driving clips as inputs and applies 12 types of synthetic degradation at five severity levels. Some degradation methods use physics-inspired proxies. The dataset includes: 10 driving scenarios 610 video clips and 78,080 frames A configured, severity-derived Global Sensor Health Index reference (gshi_gt) and PerceptionHealthNet estimates (gshi_pred) DRIVE-C supports controlled evaluation of perception models under synthetic degradations. Version 1.1 corrects metadata and documentation and updates the PerceptionHealthNet baseline predictions used in the revised manuscript. The video frames and generated corruption variants are unchanged from Version 1.0. Scenario S07 is identified as a nighttime freeway scene. gshi_gt is derived from generator settings and declared taxonomy weights; it is not a physical sensor measurement or a measure of downstream safety or accuracy. gshi_pred contains model estimates and should be treated as baseline outputs, not validated health labels. DRIVE-C is a compact testbed, not a large-scale standardized benchmark or a substitute for automotive-grade sensor validation. Code and pipeline: https://github.com/shiv-aher/drive-c-dataset