README.md

---
license: cc-by-nc-4.0
task_categories:
  - other
tags:
  - eye-tracking
  - pupil-segmentation
  - ambient-light
  - infrared
  - pupil
  - wearable
size_categories:
  - 1M<n<10M
viewer: false
---

# AmbientEye Dataset

**AmbientEye** is a large-scale eye-tracking dataset(n = 2,606,225) collected exclusively outdoors under natural ambient sunlight — without any active infrared illuminator. It is designed to benchmark pupil segmentation methods under varying, uncontrolled NIR irradiance conditions that arise in real-world outdoor use. 

To obtain high-quality pupil annotations, a single point is manually placed within the pupil region of the first frame of each session and provided as a prompt for SAM2. In the second stage, human annotators review and refine every frame of the segmentation result to validate its accuracy. When the predicted mask does not align well with the pupil boundary, annotators correct the annotation by manually marking pupil boundary points to fit an ellipse to the pupil. 

In total, pupil annotations were obtained for 2,518,693 out of 2,606,225 frames (96.6\%).

---

## Overview

| Property | Value |
|---|---|
| Participants | 35 |
| Sessions | 70 (2 per participant) |
| Total frames | 2,606,225 |
| Conditions | `sunfacing`, `sunoccluded` |
| Camera resolution | 400 × 400 px (monochrome IR) |
| Frame rate | 120 fps |
| Cameras per session | 2 (medial `eye0`, lateral `eye1`) |
| Calibration trials | 80 per session |
| IR illumination | **None** — ambient sunlight only |
| Recording device | Xreal Air 2 Glasses (with OV6211 IR cameras) |
| Recording period | April 2026 |

---

## Conditions

Each participant completed two sessions in different lighting orientations:

| Condition | Description | IR irradiance (mean ± range) |
|---|---|---|
| `sunfacing` | Participant faces toward the sun (higher NIR) | 220 µW/cm² (9.8 – 412.6) |
| `sunoccluded` | Participant faces away from the sun (lower NIR) | 66.5 µW/cm² (5.9 – 112.3) |

The large irradiance range across participants reflects real-world variability in sun angle, cloud cover, and participant orientation.

---

## Hardware

Recording was performed using **Xreal Air 2 Glasses** equipped with two embedded **OV6211** monochrome IR cameras:

- `eye0` — medial camera (nose side)
- `eye1` — lateral camera (ear side)

Both cameras operate at 400 × 400 px and 120 fps. No active IR LED is present; all NIR irradiance originates entirely from ambient sunlight.

---

## Participants

35 participants (19 male, 16 female).

| Ethnicity | Count |
|---|---|
| Asian | 20 |
| White | 8 |
| Black or African American | 4 |
| Middle Eastern or North African | 1 |
| Hispanic, Latino, or Spanish | 1 |
| Eastern African | 1 |

3 participants wore makeup. Full demographics are in `participant.csv`.

---

## Dataset Structure

```
AmbientEye/
├── video/                              # Raw IR video files
│   └── {PID}_{condition}/              # e.g. P1_sunfacing, P1_sunoccluded
│       ├── eye0.mp4                    # Medial camera (nose side), 400×400 px, 120 fps
│       └── eye1.mp4                    # Lateral camera (ear side), 400×400 px, 120 fps
├── frames/                             # Extracted JPEG frames — NOT included; generate with extract_frames.py
│   └── {PID}_{condition}/
│       ├── eye0/
│       │   ├── frame_00000.jpg
│       │   └── ...
│       └── eye1/
│           ├── frame_00000.jpg
│           └── ...
├── jsons/                              # Human-reviewed SAM2 segmentation contours
│   └── {PID}_{condition}/
│       ├── eye0_contours_reviewed.json
│       └── eye1_contours_reviewed.json
├── meta/                               # Session timing and calibration metadata
│   └── {PID}_{condition}/
│       ├── calibration_log.json        # 80-trial stimulus targets + timestamps
│       └── info.player.json            # Recording timing (duration, start times)
├── sample_contour/                     # Sample videos with contour overlays
│   └── {PID}_{condition}/
│       ├── eye0.mp4
│       └── eye1.mp4
├── participant.csv                     # Participant demographics (N=35)
├── IRdata.csv                          # Per-session NIR irradiance measurements
├── session_recording_info.csv          # Per-session date, time, weather, temperature
├── solar_position_reference.md         # Solar azimuth/altitude tables per recording date
└── extract_frames.py                   # Script to extract frames from video into frames/
```

---

## File Descriptions

### `video/{PID}_{condition}/eye0.mp4` / `eye1.mp4`
400×400 px monochrome IR video at 120 fps from two OV6211 cameras embedded in the Xreal Air 2 Glasses.
- `eye0` — medial camera (nose side)
- `eye1` — lateral camera (ear side)

Captured under natural ambient illumination only. No active IR LED is used.

### `frames/{PID}_{condition}/eye{0,1}/frame_XXXXX.jpg`
Individual frames extracted from the corresponding `eye0.mp4` / `eye1.mp4` video, zero-padded five-digit frame index. Provided for workflows that do not require video decoding.

**This directory is not included.** Run `extract_frames.py` to generate it (see [Frame Extraction](#frame-extraction) below).

### `jsons/{PID}_{condition}/eye0_contours_reviewed.json` / `eye1_contours_reviewed.json`
SAM2 pupil segmentation contours, reviewed and corrected by human annotators. Each file contains:
```json
{
  "video": "eye0.mp4",
  "fps": 120.0,
  "width": 400,
  "height": 400,
  "frames": [
    {
      "frame": 0,
      "contours": [
        { "points": [[x, y], ...], "area": 1234.5 }
      ]
    },
    ...
  ]
}
```

### `meta/{PID}_{condition}/calibration_log.json`
80 stimulus trials per session. Each entry records the screen position `(x, y)` of the shrinking-circle stimulus and the Pupil-synchronized timestamp of the participant's keypress response. Used for aligning eye video frames to known viewing directions.
```json
{
  "trials": [
    { "trial": 1, "pos": [683, 606], "click_ts": 165097.33 },
    ...
  ]
}
```

### `meta/{PID}_{condition}/info.player.json`
Recording timing fields only:
```json
{
  "duration_s": 171.79,
  "start_time_synced_s": 165082.55,
  "start_time_system_s": 1776402507.93
}
```
- `start_time_synced_s` — Pupil-synchronized clock at recording start
- `start_time_system_s` — Unix timestamp at recording start
- `duration_s` — total recording duration in seconds

### `participant.csv`
Participant demographics: ID (P1–P35), gender, age, ethnicity, country, makeup.

### `IRdata.csv`
Measured NIR irradiance (µW/cm²) per participant for each condition (`sunfacing_ir`, `sunoccluded_ir`), confirming ambient light levels at recording time.

### `session_recording_info.csv`
Per-session metadata: recording date, anonymized start/end times, duration (minutes), weather condition, and temperature (°C / °F).

### `solar_position_reference.md`
Hourly solar azimuth and altitude tables for each recording date, with per-session estimates linearly interpolated to each session's start time. 

### `sample_contour/{PID}_{condition}/eye{0,1}.mp4`
Short sample videos with SAM2 contour overlays rendered on the raw IR frames, provided for quick visual quality inspection.

### `extract_frames.py`
Script to extract all video frames from `video/` into `frames/` as JPEG files. 

---

## Frame Extraction

The `frames/` directory is not distributed with the dataset. To generate it:

```bash
pip install opencv-python
python extract_frames.py
```

The script reads from `VIDEO_DIR` and writes to `FRAMES_DIR` (imported from a `config` module). Create a `config.py` in the same directory with paths appropriate for your environment:

```python
# config.py
from pathlib import Path

VIDEO_DIR  = Path("video")
FRAMES_DIR = Path("frames")
```

---

## Usage

```python
import json, cv2

session = "P1_sunfacing"

# Load contours
with open(f"jsons/{session}/eye0_contours_reviewed.json") as f:
    contours = json.load(f)

# Load video
cap = cv2.VideoCapture(f"video/{session}/eye0.mp4")

for frame_data in contours["frames"]:
    cap.set(cv2.CAP_PROP_POS_FRAMES, frame_data["frame"])
    ret, img = cap.read()
    if not ret:
        break
    for c in frame_data["contours"]:
        pts = [[int(p[0]), int(p[1])] for p in c["points"]]
        # draw / process contour ...

cap.release()
```

---

## License

This dataset is released under the [Creative Commons Attribution Non-Commercial 4.0 (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/) license.

You are free to share and adapt the material for non-commercial purposes, provided appropriate credit is given.

---

## Citation

If you use AmbientEye in your research, please cite:

```bibtex
@dataset{ambienteye2026,
  title   = {AmbientEye: A Dataset for Pupil Segmentation under Natural Ambient Infrared Illumination},
  year    = {2026},
  url     = {https://huggingface.co/datasets/migHug/AmbientEye},
  license = {CC BY-NC 4.0}
}
```