Best Practices & Troubleshooting
Essential tips for creating effective models and solving common issues.
📊 Data Collection
Diversity is Key
Capture images representing the full range of conditions:
| Priority | What to Capture | Why |
|---|---|---|
| 🎯 High | Various angles (high, low, side) | Different robot viewpoints |
| 🎯 High | Multiple distances (near, medium, far) | Detection at all ranges |
| 🎯 High | Different orientations | Objects appear rotated |
| 🎯 High | Partial occlusions | Real-world scenarios |
| 🎯 High | Background variations | Different field locations |
| ⚠️ Low | Lighting conditions | YOLO-Pro handles this well* |
*Only capture specific lighting if experiencing actual detection failures.
Quality Over Quantity
100 diverse images > 500 similar images
Dataset Structure
✅ Good Dataset:
├── Training (80%)
│ ├── Multiple angles & distances
│ ├── Various orientations
│ └── Different conditions
└── Testing (20%)
└── Never seen during training
❌ Bad Dataset:
├── Training (80%)
│ └── All similar images
└── Testing (20%)
└── Images from training set
🎓 Model Training
Start Simple
- Use default parameters
- Train for 100 epochs first
- Only increase complexity if needed
Why? Simple models train faster and are easier to debug.
Monitor Overfitting
| Metric | Training | Testing | Status |
|---|---|---|---|
| Accuracy | 95%+ | <70% | ❌ Overfitting |
| Accuracy | 85% | 82% | ✅ Good |
Solutions:
- Add more diverse data
- Enable data augmentation
- Reduce training epochs
Data Augmentation Guide
| Augmentation | Setting | Use When |
|---|---|---|
| Brightness | ±20% | Always recommended |
| Rotation | ±15° | Objects can tilt |
| Zoom | 90-110% | Distance varies |
| Flip | Horizontal/Vertical | Objects can flip |
Don't Over-Augment
Keep variations realistic. A cone won't appear upside down! (you may not want to pick it up anyways if it is ;))
Training Parameters
Epochs:
├── Start: 100
├── Good performance: Stop
├── Underfitting: 200-300
└── Overfitting: 50-75
Learning Rate:
├── Default: 0.001 (usually optimal)
├── Unstable: 0.0001
└── Too slow: 0.005 (careful!)
🔧 Troubleshooting
| Issue | Symptoms | Solutions |
|---|---|---|
| Low Accuracy (<70%) | Many missed detections, poor test performance | • Add diverse training images (focus on failures) • Check label quality and consistency • Increase training epochs (200-300) • Check image quality (blur, focus, lighting) |
| False Positives | Detections where no objects exist | • Increase confidence threshold (0.5 → 0.6) • Add negative examples (images without objects) • Tighten bounding boxes in training • Check for labeling inconsistencies |
| Missing Objects | Objects clearly visible but not detected | • Lower confidence threshold (0.5 → 0.3) • Check if objects are too small • Verify camera settings (focus, exposure) • Add more examples of missed object types |
📚 Resources
Official Documentation
Tools
- Netron - Visualize models
💡 Final Thoughts
Success in FIRST Robotics AI:
- 🎯 Systematic Approach - Follow best practices consistently
- 🔄 Continuous Iteration - Each competition improves your model
Remember
The goal isn't perfection—it's continuous improvement.
Each iteration makes your system better. Each competition provides valuable data for the next version.
Good luck, and may your detections be accurate and your inference be fast! 🤖