Modelling and Simulation of Neuromorphic Datasets for Anomaly Detection in Computer Vision
A new Unity-based simulator creates custom datasets for event-based cameras, solving a major data bottleneck.
A research team from multiple institutions, including Martin Trefzer, Mike Middleton, and seven others, has published a paper introducing ANTShapes (Anomalous Neuromorphic Tool for Shapes), a novel framework designed to solve a fundamental data scarcity problem in neuromorphic computer vision. The core challenge is that physical Dynamic Vision Sensors (DVS)—event-based cameras that mimic biological vision—are expensive and limited, resulting in a severe shortage of diverse, real-world datasets for training AI models. ANTShapes directly addresses this by providing a configurable simulator to generate synthetic, event-based video data on demand, enabling research in areas like anomaly detection and object localization that were previously constrained by data availability.
The tool is built on the Unity game engine and simulates abstract 3D scenes populated by objects with randomly generated behaviors, such as specific motion and rotation patterns. A key innovation is its use of statistical processes based on the central limit theorem to automatically sample these behaviors and label objects acting anomalously. Researchers can generate datasets of arbitrary size by adjusting a limited set of parameters, exporting both the simulated event streams and corresponding frame-by-frame labels. This move towards high-quality synthetic data generation represents a significant shift for the neuromorphic vision community, potentially accelerating model development by providing tailored, scalable training environments for specialized AI applications beyond the reach of traditional camera datasets.
- ANTShapes is a Unity-based simulator that generates synthetic datasets for neuromorphic (event-based) computer vision, addressing a critical lack of real Dynamic Vision Sensor (DVS) data.
- The framework creates configurable 3D scenes where object behaviors (motion, rotation) are randomly generated and anomalies are statistically labeled, allowing for unlimited, bespoke dataset creation.
- It enables research in object recognition, localization, and anomaly detection by providing exportable event streams and labels, accelerating AI model development for specialized vision tasks.
Why It Matters
It removes a major data bottleneck for neuromorphic AI research, allowing faster development of vision systems for robotics, autonomous vehicles, and industrial monitoring.