SAED:Structure-Aware Event Dataset

SARE (Structure-Aware Reliability Evaluation Dataset), accepted for publication at the International Joint Conference on Neural Networks 2026, is a structure-centric event-based dataset designed for analyzing spatio–temporal structural reliability in neuromorphic vision systems.

Unlike conventional task-driven benchmarks, SARE focuses on controlled structural variation rather than semantic diversity, enabling systematic study of how structural degradation affects neural inference. The dataset is built using geometry-driven objects with clear edges and contours, allowing interpretable evaluation of structural integrity.
It contains approximately 27,000 event clips captured using a DVXplorer event camera (640×480 resolution) under controlled indoor conditions. The dataset is organized across object geometry, spatial scale (0.5×, 1.0×, 1.5×), and motion conditions (object motion and camera motion), providing diverse yet comparable spatio–temporal event structures.
A key feature of SARE is the inclusion of standardized structural degradation protocols, including noise injection, over-denoising, and localized structural removal. These perturbations are designed to decouple structural corruption from global event statistics, enabling reliable diagnosis of silent failure in both spiking and non-spiking neural models.
SARE is intended for reliability analysis, robustness evaluation, structure-aware learning, and the development of diagnostic metrics for event-based vision.