🛡️ AI Safety Object Detector

MAXIMUM ACCURACY MODE
Advanced YOLOv8 with Enhanced NMS & False Positive Suppression

📸 Single Image Detection

Upload Image for Detection

Core Parameters - Adjust for optimal results

🎯 Confidence Threshold

Higher = fewer but more accurate detections (recommended: 0.25-0.45)

0.05 0.95

📦 IoU Threshold

Higher = less overlap filtering (recommended: 0.45-0.55)

0.1 0.95

Performance Enhancers - Enable for maximum accuracy

✨ Test-Time Augmentation (TTA)

Multiple augmented predictions (+3-7% mAP, slower)

🎭 Multi-Scale Ensemble

Multiple image sizes (+2-5% mAP, much slower)

🎨 Image Enhancement

Auto contrast, sharpness & brightness boost

📐 Input Image Size

Higher = better for small objects (slower)

📁 Batch Processing Mode

Upload Multiple Images (JPG, PNG)

💡 Tip: Upload multiple images to process them all at once and download as ZIP

🎨 Detection Results

Annotated Image with Detections

📤 Upload an image to start detecting objects

All Detections Sorted by Confidence

📦 Batch Processing Results

📊 Batch Statistics & Details

📥 Download All Predictions (ZIP)

🎯 Recommended Settings by Use Case

🏆 MAXIMUM ACCURACY (Best for Critical Applications)

Perfect for: Safety inspections, compliance checks, detailed analysis

Parameter	Value	Why?
Confidence	`0.35-0.45`	Filters out most false positives while keeping real objects
IoU	`0.45-0.55`	Good balance for overlapping objects
TTA	✅ Enabled	+3-7% accuracy through augmentation
Ensemble	✅ Enabled	+2-5% accuracy through multi-scale detection
Enhancement	✅ Enabled	Improves detection on low-quality images
Image Size	`800-1024px`	Better for small and distant objects

Expected Performance: Best accuracy, ~5-10 seconds per image

⚡ BALANCED MODE (Speed + Accuracy)

Perfect for: General use, moderate batch processing

Parameter	Value	Why?
Confidence	`0.30-0.40`	Good detection rate with acceptable false positives
IoU	`0.45-0.50`	Standard NMS threshold
TTA	✅ Enabled	Worth the small speed cost
Ensemble	❌ Disabled	Too slow for marginal gains
Enhancement	✅ Enabled	Fast and helpful
Image Size	`640px`	Fast and sufficient for most cases

Expected Performance: Good accuracy, ~2-3 seconds per image

🚀 SPEED MODE (Real-time/Batch)

Perfect for: Large batches, real-time monitoring, quick scans

Parameter	Value	Why?
Confidence	`0.40-0.55`	Higher threshold = fewer detections but faster
IoU	`0.50-0.60`	Standard NMS, less computation
TTA	❌ Disabled	Too slow for speed mode
Ensemble	❌ Disabled	Significantly slower
Enhancement	❌ Disabled	Save preprocessing time
Image Size	`640px`	Fastest inference size

Expected Performance: Fast, ~0.5-1 second per image

🔍 Understanding Each Parameter

Confidence Threshold (0.05-0.95)

What it does: Minimum probability score for a detection to be kept
Lower (0.15-0.25): More detections, more false positives
Higher (0.40-0.60): Fewer detections, fewer false positives
Sweet spot: 0.30-0.40 for most use cases

IoU Threshold (0.10-0.95)

What it does: Controls how much boxes can overlap before one is removed (Non-Maximum Suppression)
Lower (0.30-0.40): More aggressive overlap removal, fewer boxes kept
Higher (0.50-0.70): Keeps more overlapping boxes (good for crowded scenes)
Sweet spot: 0.45-0.55 for most use cases

📚 Quick Start Examples

Try these configurations for common scenarios:

Single clear object (like fire extinguisher):
- Confidence: 0.40, IoU: 0.50, TTA: ✅, Ensemble: ❌, Size: 640px
Multiple small objects:
- Confidence: 0.25, IoU: 0.45, TTA: ✅, Ensemble: ✅, Size: 1024px
Fast batch processing:
- Confidence: 0.45, IoU: 0.55, TTA: ❌, Ensemble: ❌, Size: 640px
Low quality/dark images:
- Confidence: 0.30, IoU: 0.50, TTA: ✅, Enhancement: ✅, Size: 800px

🚀 Built with Innovation & Passion

24-Hour National Hackathon • 3-Month Innovation Journey • Student-Led Excellence

⚠️ AI-Powered Tool • Always verify critical detections manually
Made with ❤️ for safety and security applications

⚡ TechXNinjas