Piano Finger-Tracking Fairness Benchmark: Tabular Ground-Truth, MIDI, and Pose-Accuracy Results (No Video)

This dataset is the non-video portion of a benchmark for evaluating vision-based hand-pose / fingertip estimation on a piano keyboard, with a focus on fairness across skin tone, lighting, and hand size. It accompanies the project "Comparative Evaluation of Vision-Based Finger Tracking for Piano Interaction." 38 participants each played predefined key patterns on a digital piano under an overhead camera. For every pressed key (a MIDI note-on), the video frame nearest the event was passed through three pose pipelines (MediaPipe Hands, OpenPose Hand, and a per-hand cropped OpenPose variant), and the horizontal pixel distance between the detected fingertip and the calibrated key centre was recorded. This release contains the ground truth, the participant covariates, and the resulting accuracy metrics. Important: this release contains NO video and NO per-frame landmarks. Fingertip positions were estimated during analysis and were not stored per frame; the source video is not published (only hand regions were ever recorded, and video is excluded to protect privacy). What is published is sufficient to reproduce the study's fairness analysis and to reuse the MIDI and key-calibration ground truth, but not to re-run pose estimation. Contents (5 CSV files):- participants.csv (38 rows) — per-participant covariates: Fitzpatrick skin type, ambient lux, hand size, video fps.- midi_events.csv (9,557 rows) — MIDI note on/off events with timestamps and pitch names.- key_calibration.csv (912 rows) — piano-key pixel-centre coordinates per participant.- results_summary.csv (114 rows) — accuracy summary per participant × model.- results_per_finger.csv (1,140 rows) — error and accuracy per participant × model × hand × finger. Covariate spread: Fitzpatrick skin types II (4), III (16), IV (13), V (5); lighting Dim (28), Indoor (10); hand size 7.7–12.2 cm. Documentation included: README.md (overview, quick-start, join keys), SCHEMA.md (column-by-column data dictionary), DATASHEET.md (following Datasheets for Datasets, Gebru et al. 2021), CITATION.cff, and LICENSE. Ethics & privacy: Data was collected under signed participant consent with Leeds Beckett University research-ethics approval. Only hand regions were recorded; all records are de-identified (participant ID codes only — no names, faces, or video).