thermophysio: FLIR thermal psychophysiology dataset

This dataset accompanies the thermophysio Python library for extracting and analysing skin temperature time-series from FLIR radiometric thermal infrared camera recordings of human participants during psychophysiology experiments.

Experiments: Twenty-four sessions across five cognitive task types are included:

Faces (Exp01–05, Exp08) — passive face viewing; back-of-head camera

GoNoGo (Exp06, Exp09, Exp11) — response inhibition; frontal camera

Words (Exp07, Exp10, Exp12, Exp21) — lexical decision; frontal camera

CORSI (Exp13–20) — visuospatial working memory; frontal camera

LogicalVsIllogical (Exp22–24) — logical reasoning; dual side-profile cameras

Camera configurations Three FLIR camera setups were used: frontal face-forward (MediaPipe 468-landmark ROI extraction, 13 regions), back-of-head (YOLOv8 ear-keypoint basis, 11 regions), and dual side-profile (YOLOv8 oblique eye–ear–nose basis, 24 bilateral regions).

Archive contents

Experiments.zip — raw FLIR radiometric JPEG images, one folder per session

Processed_Experiments.zip — extracted pixel arrays (.npz), session metadata (JSON), and ROI overlay images; output of the interactive extraction step

Calc_Experiments.zip — per-session analysis outputs: 10-feature statistics CSVs, resampled time-series, Bartlett-corrected cross-correlation tables, leaderboard plots, and Fisher z meta-analysis reports

Usage

The dataset is designed to be used with the thermophysio library (https://github.com/RauwP/thermophysio). Raw images in Experiments/ can be re-processed using the library's extraction pipeline. Pre-extracted data in Processed_Experiments/ can be fed directly into the analysis pipeline without repeating the interactive ROI placement step.

Analysis method

For each session, nine thermal features (mean, median, max, min, variance, energy, Shannon entropy, kurtosis, skewness) are extracted per anatomical ROI per timepoint. Pairwise differences between regions are cross-correlated with a task activation reference signal using Bartlett-corrected confidence intervals to account for autocorrelation. Cross-session consistency is assessed via Fisher r-to-z meta-analysis with between-run confidence intervals.