Event-Based Tiny Images dataset

The event-based tiny image datasets were created by capturing popular image classification benchmarks — MNIST, Fashion-MNIST, and Kuzushiji-MNIST — using a DAVIS240C event-based camera (iniVation AG) mounted on a robotic pan-tilt unit. Each source dataset's 28×28 pixel symbols were displayed on a computer screen and recorded at two camera distances: 525 mm (D1) and 830 mm (D2), resulting in projected symbol sizes of approximately 5×5 and 4×4 sensor pixels, respectively, with display-to-sensor pixel ratios of roughly 1:7–1:8 and 1:12–1:13. The camera was moved horizontally (and occasionally vertically) at a nominal speed of ~30 deg/s to generate luminance-change events, producing asynchronous spatio-temporal streams in which each event is encoded as a tuple of timestamp, x-y sensor coordinates, and polarity. The resulting six event-based datasets (named with the "ebt" prefix and a distance suffix, e.g., ebtMNIST-D1) each contain 60,000 training and 10,000 test samples, matching the size of their source datasets. Equivalent frame-based counterparts ("fbt" prefix) were also collected with the camera stationary at a 20 Hz frame rate and 6 ms exposure, serving as spatial-only baselines for comparison.