A Spinning Teapot Teaches Machines to See in the Dark

A cube turns slowly in an empty black scene. No camera shutter clicks. No frame captures a still image at thirty times a second. Instead, an event camera records local changes in brightness. Light and simulated surface texture shift across the rotating cube. Its pixels emit a stream of timestamped events. That stream forms a distinctive spatiotemporal signature.

Somewhere nearby, a Utah teapot spins through the same darkness. Its curved handle traces a different rhythm of flickers entirely. Neither object is real. Researchers generated both in simulation to solve a specific bottleneck in event-based artificial intelligence. The field has had few purpose-built benchmarks for this task. Those benchmarks would test whether a neuromorphic system can classify three-dimensional objects. The test relies on evolving event signatures alone.

That gap is the starting point for new research out of the University of York, Cardiff University, the University of Manchester, the University of Stirling and the University of Southampton. M. Middleton and colleagues lead the work. It introduces ANTShapes, short for Anomalous Neuromorphic Tool for Shapes. ANTShapes is a simulation environment. It renders moving three-dimensional objects and exports their brightness change events. The output uses an Address Event Representation format, the same encoding style physical event cameras use. The paper presents four such synthetic datasets. It benchmarks them against the datasets most commonly borrowed for this task today.

Three Technologies, Not One

Two related technologies sit behind this research, and it helps to separate them early. Event cameras are sensors that do not capture complete images at fixed intervals. Instead, each pixel independently emits an event whenever it detects a sufficient local change in brightness. That event records coordinates and a timestamp. It also records whether the change was an increase or decrease in light.

Spiking Neural Networks, or SNNs, are algorithms inspired by biological neurons. They suit the task of processing these sparse, asynchronous event streams well. Neuromorphic chips are specialized hardware built to run networks like these efficiently. Examples include Intel’s Loihi, the University of Manchester’s SpiNNaker platform and the BrainScaleS system. The three technologies often appear together in a single pipeline. Even so, researchers can pair each one with other hardware or algorithms.

Why the Event-Driven Approach Matters

Computation in this event-driven approach happens in response to actual changes in the scene. It does not process a fixed-rate stream of full frames. That difference lets neuromorphic systems cut a good deal of unnecessary processing. Conventional frame-based pipelines carry that overhead by design. Research groups working with neuromorphic hardware have reported inference on the order of microseconds in specific configurations. Reported gains also include a substantially reduced power budget relative to conventional deep learning hardware. The exact gains depend heavily on the chip, the task and how the surrounding system is built.

That distinction still carries real weight for constrained applications. Size, weight and power budgets matter most for drones, wearable sensors, satellites and covert monitoring equipment. All of these stand to benefit from hardware built for low-latency inference at a reduced energy draw. Running inference on the device itself can also cut how much raw visual data needs to travel to a cloud server. That local processing may improve privacy and lower latency. It can also reduce dependence on a constant network connection. Even so, a deployed system may still store, transmit or log some event data or alerts, depending on how engineers design it.

Object classification sits at the center of nearly every downstream neuromorphic vision task. Scene understanding depends on it. So does autonomous navigation for robots and vehicles, along with anomaly detection in industrial or security settings. Each of these tasks depends on a system first correctly identifying what it is looking at. The quality of the training and evaluation datasets therefore has an outsized influence on the whole pipeline.

The Trouble With Borrowed Data

To understand what ANTShapes sets out to correct, the authors first examined four datasets. These have become the default choices in neuromorphic vision research, namely N-MNIST, CIFAR10-DVS, POKER-DVS and DVSGesture.

N-MNIST takes the familiar handwritten digit dataset and converts it into spike form. Researchers mounted an event camera on a moving platform. They recorded it as the camera performed small triangular sweeps in front of a screen. CIFAR10-DVS takes a similar approach. It moves static images of animals and vehicles across a monitor while a Dynamic Vision Sensor camera records the resulting flicker.

Both datasets produce genuine spike streams. Yet the motion generating those streams comes entirely from the recording apparatus itself. Nothing in the scene actually moves on its own. A digit does not evolve. An image of a horse does not shift its pose. The spatial content stays fixed while only the sensor’s viewpoint changes. These datasets therefore carry little of the natural temporal richness a real object classification task would eventually require.

Two Datasets Rooted in Real Scenes

POKER-DVS and DVSGesture solve part of that problem. Both capture genuinely dynamic, real world scenes with an actual event camera. POKER-DVS records playing cards flicking past a sensor at high speed. DVSGesture captures twenty-nine human subjects performing eleven different hand and arm gestures. Both carry true spatiotemporal structure.

POKER-DVS distinguishes suit symbols printed on moving cards. DVSGesture distinguishes classes of human movement instead. Both contain genuine temporal richness. Neither one tests quite the same problem as identifying among multiple discrete three-dimensional object categories. On top of this, the physical recordings behind all four established datasets carry sensor noise and recording artifacts typical of DVS hardware. The authors filter that noise out during preprocessing before training on it.

Building a Cleaner World

Rather than recording more physical footage, the research group turned to simulation. The ANTShapes tool generates twelve distinct three-dimensional object classes. These span flat faced shapes like cuboids and pyramids. They also include smooth faced shapes like spheres, cones and toruses. A final group covers complex reference meshes, including the Utah teapot and the Suzanne monkey head model familiar to anyone who has used Blender.

The baseline dataset, called Standard, places a single object in an empty scene. Each object receives a randomized starting position and initial rotation. It then spins continuously around all three axes at a constant rate. Three further datasets each layer on an additional source of variation. The Translation dataset lets objects drift and bounce across the scene rather than staying anchored near the center. The Distortion dataset stretches and squashes each object along independent axes. No two examples of the same class end up sharing identical proportions. The Rotation dataset swaps the fixed rotation vector for a randomized one per object. No two spinning cubes ever complete their turn along quite the same path.

The Real Advantage: Experimental Control

The deeper advantage of this approach is experimental control. The authors define ANTShapes parameters as normal distributions with a set mean and variance. That design lets them hold the object category fixed. Meanwhile they can independently vary translation, shape proportions or rotational dynamics. It turns the benchmark into a diagnostic tool rather than a pure accuracy exercise. A researcher can ask which specific kind of variation causes a classifier to fail. That question matters more than a single overall performance score. In a physical recording, lighting, sensor noise, viewpoint, background texture and motion tend to change together. That entanglement makes the same diagnostic question far harder to isolate.

Each dataset contains three thousand examples, roughly two hundred and fifty per class. An eighty-twenty split produces twenty four hundred training samples and six hundred test samples. Every recording lasts eight hundred thirty six milliseconds. The temporal resolution runs at one millisecond per event. A shape completes a full rotation on its axes roughly every two hundred nine milliseconds, using the unit rotation vector from the Standard configuration.

Objects also carry a normal mapped surface texture during rendering. This lighting dependent visual noise differs from the sensor noise stripped from the physical DVS datasets. The entire environment is generated computationally, so the authors could remove that sensor level noise entirely. They still retained fine control over how much genuine variation each class experiences. That combination gives researchers a parameterized testing ground. Neither N-MNIST nor CIFAR10-DVS nor any of the natively recorded datasets could offer that on their own.

Putting the Data to the Test

To check whether these synthetic scenes behaved like useful training material, the authors built a convolutional spiking neural network using the snnTorch library. They trained it separately on all eight datasets, the four established benchmarks and the four new ANTShapes variants. Each network used leaky integrate and fire neurons. Surrogate gradient training worked around the non differentiable nature of spike emission. A final non spiking readout layer translated accumulated membrane potentials into a class prediction.

The number of hidden convolutional layers scaled with each dataset’s spatial resolution. Lower resolution inputs such as N-MNIST and POKER-DVS used a shallower network. Higher resolution datasets like DVSGesture and CIFAR10-DVS needed more layers. The researchers trained ten independently initialized networks per dataset. That approach captured how consistent the results were across different random starting conditions.

How the Datasets Compared

The results lined up closely with expectations. The Standard ANTShapes dataset reached 92.2 percent mean classification accuracy, the highest of the four synthetic variants. It represents the simplest case. No added translation, distortion or rotational randomness applies beyond the randomized starting orientation every example already includes. The Distortion dataset came in lowest among the synthetic sets at 83.5 percent. That result reflects the genuine added difficulty of classifying objects whose proportions vary from one example to the next. Translation and Rotation landed in between at 89.1 and 89.8 percent respectively.

These figures sat comfortably alongside the established benchmarks. N-MNIST and POKER-DVS reached 97.0 and 96.6 percent mean accuracy respectively. Those scores are consistent with the relative separability of their classes under this experimental setup. Still, the comparison is not perfectly apples to apples. The eight datasets differ in class count, resolution, sample count and network depth. DVSGesture reached 88.9 percent.

CIFAR10-DVS, by contrast, lagged far behind at just 55.6 percent. The authors trace that weaker result to two factors. One is the dataset’s strong imbalance toward positive polarity events. The other is substantial class overlap visible in their principal component analysis. Those diagnostics point toward a likely cause without settling the question entirely. Confusion matrices generated across all eight datasets showed a clear diagonal for every ANTShapes variant. The classifier recovered substantial class specific structure. Most errors concentrated among visually similar forms, such as the Tetris style L-block and T-block objects.

Reading the Confusion

Principal component clustering added a further layer of confirmation. Sphere samples were among the most clearly separated classes across every ANTShapes variant. Their outer silhouette remains nearly invariant under rotation. The simulated surface texture and constant lighting still produce a comparatively consistent pattern of internal events as the sphere turns. More angular objects present substantially different edges and faces at different orientations. That difference made their signatures more variable. Some pairs, such as the L-block and T-block objects, became more easily confused with one another as a result.

Where CIFAR10-DVS showed extreme overlap between nearly all ten of its classes on the PCA plot, the ANTShapes datasets retained clearer boundaries. That clarity held even under the added stress of translation, distortion or randomized rotation. The contrast supports one of the paper’s central claims. Datasets built from converted static images may lack the separable structure a fair object classification test needs, no matter how widely the literature uses them.

What the Authors Say About Scope

The authors are candid about the boundaries of a single benchmarking pass. Researchers tested only one convolutional SNN architecture. The gap in accuracy between the easiest and hardest ANTShapes variants stayed under ten percentage points. That narrow gap suggests future versions of the tool may need sharper difficulty gradients. Sharper gradients would help the benchmark keep pace with increasingly capable network designs. The classification task is also static. Only one object class appears per example, and that class never changes mid recording.

Additional Boundaries Worth Noting

A few additional boundaries are worth keeping in view alongside those the authors raise directly. ANTShapes is a synthetic, single object benchmark rendered against an empty background under fixed lighting. The networks in this study trained and tested within each dataset. They did not train in simulation and then face evaluation against recordings from a physical event camera. The paper therefore does not yet measure how well a classifier trained on ANTShapes would transfer to real world footage.

The input resolutions and resulting network depth also vary across the eight datasets compared here. Readers should treat the experiment as validating ANTShapes as usable, well separated training material. It is not a single, fully architecture controlled ranking across every benchmark involved.

Looking further ahead, the ANTShapes tool already supports simulating scenes with multiple objects. It can automatically label anomalous behavior using pixel level masks and bounding boxes. The authors flag this foundation as ripe for expansion. One clear direction is scene decomposition research. A system there would need to isolate individual objects within a busier, multi object environment. Only then could it classify each one.

From the Lab to the Field

It is worth being direct about what this paper does and does not show. ANTShapes does not yet establish performance on drones, wildlife cameras, factory floors or doorbell footage. Its contribution is more foundational than that. It gives researchers a controlled way to test whether an event based classifier can learn object specific temporal structure. That test happens before the classifier ever faces the messier conditions of a real deployment. Those conditions include clutter, occlusion, changing backgrounds, shifting distance and viewpoint, genuine sensor noise and the broader gap between simulated and real world data.

Where the Research Could Eventually Lead

With that caveat in place, the longer term motivation for this kind of research is still worth sketching out. Neuromorphic classifiers refined against richer, more controlled spatiotemporal benchmarks could eventually contribute to equipment monitoring in factories. A spiking system might watch for subtle changes in machine vibration or part alignment. It could flag a fault well before a conventional camera based system completes its next frame capture.

Environmental monitoring stations placed in remote areas could benefit too, especially where reliable power infrastructure is scarce. Such a station could run for extended periods on battery power alone. It would wake only when motion in the scene genuinely warrants attention. Search and rescue drones equipped with event based sensors could eventually help as well. They might distinguish a person from surrounding debris in low light conditions. That distinction would come at a fraction of the compute budget frame based alternatives require.

Even everyday consumer devices could benefit one day. Smart doorbells and wearable safety monitors are two examples. Classifiers trained on data engineered specifically to separate meaningful object shapes from background clutter would serve them well. That would mark a real change from data repurposed from tasks the original dataset creators never intended for object recognition at all. None of these outcomes follow directly from the results reported here. Still, they help explain why closing the benchmark gap matters in the first place.

A Cleaner Testbed, Not a Finished System

As neuromorphic hardware continues moving from research labs toward practical edge systems, benchmark design matters nearly as much as chip design. ANTShapes does not replace real world event camera data, and the authors do not claim that it does. What it supplies is a controlled complement to that data.

Letting researchers vary shape, motion, rotation and distortion while holding other conditions steady makes classifier failures considerably easier to interpret. A network that falters under distortion points toward a robustness problem worth investigating further. A network that only succeeds in a noise free simulation points toward the next experiment the field still needs to run. That experiment would involve real sensor data. That kind of diagnostic clarity may end up being the most durable contribution this line of research delivers.


Citation: Middleton, M., Kayan, H., Sen Bhattacharya, B., Ali, T., Baikas, E., Vousden, M., Perera, C., Rhodes, O., Gheorghiu, E., & Trefzer, M. A. (2026). ANTShapes Benchmarking Datasets for Event-Based Neuromorphic Object Classification. arXiv:2608.27150v1 [cs.NE]. https://arxiv.org/abs/2608.27150

Related Articles

Responses

Your email address will not be published. Required fields are marked *

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.