What four datasets does ANTShapes add to event-based neuromorphic classification?

A new paper on arXiv introduces four benchmark datasets for event-based neuromorphic object classification. The datasets come from the ANTShapes simulation tool and are tested against four established event-vision benchmarks.

Event cameras replace frames with asynchronous spikes

Event cameras do not capture frames. They report brightness changes per pixel as asynchronous events. This difference matters for devices that must stay small and light. Power consumption must also stay low. The paper submitted on 27 August 2026 describes four new datasets built for that paradigm. Each dataset carries labels ready for object classification tasks.

Frame-based pipelines hit physical and security limits

Conventional pipelines rely on synchronous frame-based cameras. Those cameras feed cloud or edge compute that consumes significant size and weight. Power draw adds a third burden. The document states this prohibits deployment at the extreme edge or in covert sensing environments. Cloud links also introduce latency and demand constant connectivity. Sensitive data must leave the device. Security concerns arise from sending and receiving potentially sensitive data. Spiking neural networks on neuromorphic hardware aim to solve these issues.

ANTShapes tool generates four new benchmark sets

The ANTShapes simulation tool was previously proposed to create and label event-based vision datasets. Research has been hindered by a shortage of high-quality event-vision datasets. Now four novel datasets of varying difficulties emerge from that tool. The paper provides rich details for future experiments. It also validates the ANTShapes simulator as fit for purpose.

Convolutional SNN tests all eight datasets together

The four new datasets join N-MNIST, CIFAR10-DVS, DVSGesture and POKER-DVS in the benchmark suite. A convolutional spiking neural network classifies objects across all eight sets. Classification is performed using a convolutional SNN. The work simultaneously provides four datasets with rich details for future experiments to use.

Four datasets. Four existing benchmarks. One convolutional SNN.

Paper metadata and open questions

The paper spans 23 pages and 11 figures. File size reaches 1,687 KB. Ten authors contribute under the Neural and Evolutionary Computing category. The DOI is 10.48550/arXiv.2608.27150. Subjects include Artificial Intelligence and Computer Vision and Pattern Recognition. What the document does not say is how many events each dataset contains. It also omits classification accuracy numbers for the new sets. Those details remain in the full text.