SadakVision Weights
Trained weights for Indian road defect + garbage detection (crack / garbage / manhole / pothole).
Dataset: Merged_Final_Data (Roboflow) — YOLO format, train/valid/test.
Files
| File | Source | Size |
|---|---|---|
yolov8m_100e_best.pt |
yolov8m_res_100/train/weights/best.pt |
49.6 MB |
yolov8m_runs_best.pt |
yolov8m_runs/pothole_yolov8m/weights/best.pt |
49.6 MB |
yolo26m_best.pt |
yolo26_results/yolo26m_runs/weights/best.pt |
42 MB |
yolov11n_best.pt |
yolov11n_res/runs_yolov11n/weights/best.pt |
5.2 MB |
rfdetr_best.pth |
rfdetr/checkpoint_best_total.pth |
127.6 MB |
Results (validation split)
| Model | mAP50 | mAP50-95 | Precision | Recall |
|---|---|---|---|---|
| YOLOv8m (100e, imgsz 640, batch 32) | 0.773 | 0.489 | 0.858 | 0.736 |
| YOLOv26m (100e, batch 16) | 0.757 | 0.386 | 0.824 | 0.737 |
| YOLOv11n (100e, imgsz 512, batch 64) | 0.633 | 0.294 | 0.721 | 0.594 |
| RF-DETR (medium) | 0.796 | 0.550 | 0.881 | 0.768 |
RF-DETR val also: F1 0.817, mAP75 0.586, mAR 0.660. Per-class AP: crack 0.730, manhole 0.686, pothole 0.525, garbage 0.259 — garbage hardest.
Test: YOLOv8m mAP50 0.742, YOLOv26m 0.741, YOLOv11n 0.644.
Usage
pip install ultralytics
yolo detect predict model=yolov8m_100e_best.pt source=your_image.jpg save=True
RF-DETR .pth is the checkpoint_best_total.pth (no optimizer state).