← Research Lab / HCI
FIG. 25A0 // FANTASMA RESEARCH NOTE

Calibration-Free Gaze Tracking at 60fps on Commodity Hardware

Oleh James Okafor
8 min baca
2026-07-03

Calibration-Free Gaze Tracking

Gaze tracking has been "almost ready" for consumer deployment for a decade. The Achilles heel: calibration. Before you get accurate estimates, users must stare at 9, 13, or 25 fixation points while a model learns their personal geometry.

Beyond the UX friction, this creates a privacy requirement — you must store per-user calibration profiles. In healthcare and defense applications, this is often untenable.

The Synthetic Data Approach

We bypassed the calibration problem by removing the need for personalization altogether.

Using a differentiable eye renderer parameterized over 180 anatomical variables (iris size, limbal ring, sclera reflectance, eyelid fold geometry, corneal curvature), we generated 4.2M unique synthetic eyes across 400,000 simulated lighting conditions. We trained a compact ViT-Tiny encoder to regress gaze direction from raw infrared frames — 30×30 pixels, 60fps, 12mW on an ARM Cortex-A55.

Accuracy

On our held-out set of 312 real users (not seen during training):

Mean angular error: 0.41° (vs. 0.38° for calibrated classical methods)
P95 error: 0.89°
Zero calibration time
No stored user data

Why This Matters

Gaze is the highest-bandwidth implicit signal a human emits — 4 saccades per second, each carrying intent. When you can read it without friction, without personalization, without a privacy footprint, the interface design space opens up completely.

Publikasi Riset Terkait

Architecture 11 min

Attention Is Not Enough: Toward Structured World Models at the Edge

Transformer attention scales quadratically with sequence length. For always-on edge inference, this is a thermodynamic impossibility. We propose a hybrid SSM-attention architecture that achieves 94% of full-attention quality at 8% of the compute.

Systems 14 min

Byzantine Consensus in Low-Power Edge Meshes: Revisiting PBFT for the Real World

Classical Byzantine fault tolerance assumes reliable networks and ample compute. Edge meshes have neither. We present a modified PBFT protocol that tolerates 30% Byzantine nodes with 2ms consensus on a 10-node mesh running at 0.8W per node.