STEMNIST: Spiking Tactile Extended MNIST Neuromorphic Dataset

STEMNIST: Spiking Tactile Extended MNIST Neuromorphic Dataset

The STEMNIST dataset has a hierarchical directory structure to support both spikebased neuromorphic processing and frame-based deep learning methods. The dataset consists of two main components: 

Raw Pressure Data: The raw 8-bit pressure data associated with each handwritten character, initially consolidated within a single five-character HDF5 file, were subsequently disaggregated and saved as individual HDF5 files through Python-based post-processing. Each sample comprises a three-dimensional array of pressure readings (240 frames × 16 × 16) accompanied by comprehensive metadata, including participant identifier, character label, repetition index, sampling frequency (120 Hz), and temporal acquisition timestamp. The file naming convention adheres to the naming structure {<ParticipantID>_<Character>_<RepetitionNumber>.h5}. 

Processed Spike Data: These are event-based data representations, stored in HDF5 files as structured NumPy arrays. Three fields are present in each spike file: timestamp (float32), taxel ID (int32) and polarity (int16). Additionally, each spike file contains attributes that display the total number of spikes, the adaptive threshold value ($\theta_{\text{sample}}^{\text{clipped}}$) and a reference to the original raw file.  Spike files use the naming convention {<ParticipantID>_<Character>_<RepetitionNumber>_spikes.h5} and are arranged into 35 subdirectories according to their respective character class.

More details about the dataset can be found here: https://arxiv.org/abs/2601.01658