eye-open-closed-dinov2

Eye Open/Closed Classifier

DINOv2-ViT-Large fine-tuned on MRL Eye Dataset (NIR grayscale eye patches).

Performance

Setup

# Install dependencies
pip install torch torchvision timm onnxruntime pillow

# Download model weights
pip install huggingface_hub
huggingface-cli download Yingzhe/eye-open-closed-dinov2 best_dinov2_eye.pth --local-dir .

Quick Start

Single Image

python eye_open_closed_infer.py --image eye.jpg

Output:

  eye.jpg: OPEN (confidence=0.9872)

Batch Mode (with labeled images)

python eye_open_closed_infer.py \
  --input_dir ./eye_images/ \
  --output_dir ./labeled_output/ \
  --draw_label \
  --conf_threshold 0.85

This will:

• Predict open/closed for every image in ./eye_images/
• Draw prediction text on each image (green=OPEN, red=CLOSED, yellow=UNCERTAIN)
• Save labeled images to ./labeled_output/
• Generate ./labeled_output/results.csv with all predictions

Custom Confidence Threshold

# Strict: only trust predictions above 0.95
python eye_open_closed_infer.py --input_dir ./eyes/ --output_dir ./out/ --draw_label --conf_threshold 0.95

# Lenient: accept all predictions
python eye_open_closed_infer.py --input_dir ./eyes/ --output_dir ./out/ --draw_label --conf_threshold 0.50

GPU Acceleration

python eye_open_closed_infer.py --input_dir ./eyes/ --output_dir ./out/ --device cuda

ONNX Inference (no PyTorch needed)

# Download ONNX model
huggingface-cli download Yingzhe/eye-open-closed-dinov2 dinov2_eye_classifier.onnx --local-dir .

# Run
pip install onnxruntime pillow
python eye_open_closed_infer.py --input_dir ./eyes/ --output_dir ./out/ --draw_label --onnx dinov2_eye_classifier.onnx

Arguments

Output CSV Format

filename,prediction,confidence
s0001_00001.png,OPEN,0.9987
s0002_00015.png,CLOSED,0.9921
s0003_00638.png,UNCERTAIN,0.5004

Label Legend

• OPEN (green) — Eyes are open, confidence ≥ threshold
• CLOSED (red) — Eyes are closed, confidence ≥ threshold
• UNCERTAIN (yellow) — Low confidence (occlusion, blur, extreme angle)

Confidence Distribution (MRL test set)

Recommendation: Use --conf_threshold 0.85 for pseudo-label generation. This filters out ~3% of uncertain samples while keeping accuracy > 0.998.

Training

See train_dinov2.py for training code. Train on MRL Eye Dataset:

• Download: kaggle datasets download -d akashshingha850/mrl-eye-dataset
• Dataset: 50,937 train / 16,980 val / 16,981 test (grayscale eye patches)
• Training time: ~15 min on RTX 4090 (20 epochs)
• Strategy: Freeze backbone 5 epochs → unfreeze → fine-tune all

Notes

• Input can be any resolution. Grayscale images are auto-converted to 3-channel.
• MRL dataset is NIR grayscale (infrared camera). Works best on similar data.
• For RGB visible-light images, performance may vary (domain gap).