ANTShapes Benchmarking Datasets for Event-Based Neuromorphic Object Classification

Event-based vision, which utilizes neuromorphic sensors to capture changes in scene intensity asynchronously, offers significant advantages in power efficiency and latency compared to traditional frame-based cameras. However, research into spiking neural networks (SNNs) for object classification is currently hampered by a lack of high-quality, noise-free, and temporally rich datasets. Existing benchmarks often rely on converting static images into event streams or focus on specific tasks like gesture recognition, which may not serve as ideal proxies for general 3D object classification.

To address these limitations, the authors utilize the ANTShapes simulation tool to generate four novel datasets. These datasets feature 12 distinct 3D object classes, including various geometric shapes and complex meshes. By using simulation rather than physical sensors, the authors eliminate stochastic sensor noise, providing a clean environment for benchmarking. The four datasets vary in difficulty:

• Standard: Baseline dataset with objects rotating around their axes.
• Translation: Objects move across the scene with random velocity.
• Distortion: Objects undergo random scaling, squashing, and stretching.
• Random Rotation: Objects rotate using independent random vectors per example.

The authors benchmark these new datasets against established standards such as N-MNIST, CIFAR10-DVS, POKER-DVS, and DVSGesture using a convolutional SNN architecture. The results indicate that the ANTShapes datasets achieve high classification accuracy, comparable to or better than existing benchmarks. The study validates that the ANTShapes tool is fit for purpose, providing a controlled and scalable way to generate data for neuromorphic vision research. The authors also note that while the datasets are effective, the classification task remains relatively simple, suggesting that future work could explore more complex, time-dependent scenarios.