IRIS Road Intelligence Model v2
YOLOv8n trained for on-device road infrastructure detection in Kenya.
15 Detection Classes
| ID | Class | Source | Use |
|---|---|---|---|
| 0 | POTHOLE | RDD2022 | Road damage assessment |
| 1 | CRACK | RDD2022 | Road damage assessment |
| 2 | PATCH | Field collection | Maintenance tracking |
| 3 | UNPAVED_ROAD | Field collection | Surface classification |
| 4 | SPEED_BUMP | Field collection | Road furniture mapping |
| 5 | ROAD_SIGN | COCO (stop sign) | Traffic management |
| 6 | TRAFFIC_LIGHT | COCO | Intersection mapping |
| 7 | GUARDRAIL | Field collection | Safety infrastructure |
| 8 | PEDESTRIAN_CROSSING | Field collection | Pedestrian safety |
| 9 | ROAD_MARKING | Field collection | Maintenance tracking |
| 10 | MANHOLE | Field collection | Utility mapping |
| 11 | DRAINAGE | Field collection | Water management |
| 12 | VEHICLE | COCO | Traffic density |
| 13 | MOTORCYCLE | COCO | Bodaboda tracking |
| 14 | CONSTRUCTION | Field collection | Roadwork zones |
| 15 | NUMBER_PLATE | COCO synthetic + ALPR | Vehicle identification (hashed) |
Training
- Model: YOLOv8n (3.2M params)
- Input: 320x320 RGB
- Dataset: 10000 images (RDD2022 + COCO subset)
- Epochs: 80
- Output: (1, 20, 2100) detection tensor
Mobile Deployment
Download iris-comprehensive-v1_float32.tflite (~12MB) for iOS/Android.