---
license: mit
library_name: pytorch
tags:
  - gaze-estimation
  - eye-tracking
  - mobilenetv2
  - webcam
---

# gazekit personal gaze models

Gaze-estimation models trained with [gazekit](https://github.com/ZoneTwelve/gazekit),
a webcam eye-tracking toolkit with a full personal training lifecycle
(calibration, VOR/posture collection scenarios, ambient background training,
clean/train/validate/evaluate/update iteration).

**These weights are personalized**: they were trained on one user's face,
one camera, and one screen. They will not work well for anyone else — treat
them as a reference artifact / starting checkpoint, and run the gazekit
pipeline to train your own.

## Files

| file | description |
|---|---|
| `gaze_cnn.pt` | MobileNetV2 backbone (ImageNet init, first conv adapted to grayscale), two 64x48 eye crops + head pose (yaw/pitch/roll) → normalized screen (x, y). See `gazekit/cnn.py` for the exact architecture. |
| `gaze_model.pkl` | Linear ridge on the 19-dim "v5-combo" features (binocular iris + distance-gain + pose interactions). **Requires gazekit at commit `f3a6261` or later** — the feature transform lives in `gazekit/model.py`, the pickle only stores the fitted pipeline. |

## Usage

```python
from gazekit.cnn import CnnPredictor
pred = CnnPredictor("gaze_cnn.pt", screen_size=(1920, 1080))
# obs comes from gazekit.tracker.FaceTracker(...).process(frame, want_crops=True)
xy = pred.predict(obs)
```

## Training data

Personal dataset (private): dwell-point calibration grids, VOR head-movement
sweeps, multi-posture grids, screen-edge points, smooth-pursuit sweeps, and
ambient popup samples, cleaned by the `gazekit iterate` pipeline.