gazekit personal gaze dataset

Personal eye-tracking dataset collected with gazekit. Contains biometric data (eye-region images) of a single individual, shared by that individual. Use for gaze-estimation research/experiments; do not use for identification or attempt to re-identify beyond the published account.

Collection toolkit

All of this data is collected by the open-source gazekit pipeline — calibration grids, VOR/posture/edges scenarios, ambient popups with a UCB sampling bandit, mouse-verify teaching, ARKit teacher pairing (ios/GazeTeacher), and a clean/train/validate/evaluate/update loop. Reproduce your own dataset with python -m gazekit auto.

Contents

dataset.tar.gz unpacks to session_*/ directories:

• samples.jsonl — one record per sample: screen-target label (px), 14-dim landmark feature vector, head pose (yaw/pitch/roll, deg), blink score, and a collection tag
• crops/NNNNNN_R.png, crops/NNNNNN_L.png — 64x48 grayscale roll-normalized eye crops (right/left)
• pruned.json at the root — sample ids flagged as noise by the gazekit iterate cleaning stage (skip these when training)

Tags

tag	scenario
calib / probe / repair	calibration grid dwell points / held-out validation probes
vor	fixed dot, moving head (vestibulo-ocular reflex)
posture	grid repeated at 3 sitting postures
edges	near-margin screen points
pursuit	smooth-pursuit sweep (labels lag-compensated; noisier)
click / ambient	live-mode click-teach / background popup samples
closed	eyes closed (blink calibration; no valid gaze label)

Load helpers: gazekit.dataset.load_sessions (CNN training pairs) and gazekit.dataset.load_dwell_features (ridge features).