eye-closure-detector-yolo-tflite

Info

• Models made from project WakeUp App
• Based on Yolov5, Yolov8.
• pt models converted to tflite to embedding on mobile device
• Trained from preprocessed dataset
• It detects two class openeyes/closedeyes

---

Data Processing phases

Three key preprocessing phases included:

1. Removing:
   • ambiguous/half-closed eyes
   • low-resolution or glare-heavy images
   • duplicates and background interference
2. Background removal & image cleaning
3. Data augmentation: brightness tuning + image flipping

Final Dataset Size: 9,200+ images

📊 Model Candidates & Results

Model	Precision	Recall	mAP@50	mAP@50–95
YOLOv5n	0.9676	0.9619	0.9631	0.4822
YOLOv5s	0.9710	0.9651	0.9632	0.4978
YOLOv8n	0.9672	0.9461	0.9629	0.5126

💡 Despite better accuracy from larger models, YOLOv5n was chosen due to its significantly better latency for mobile environments.

🔗 WANDB Logs:

• YOLOv5n Report
• YOLOv8n Report
• YOLOv5s Report