File size: 4,400 Bytes
448b62f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | ---
license: agpl-3.0
library_name: ultralytics
pipeline_tag: object-detection
tags:
- yolo
- yolo26
- road-damage
- pothole-detection
- crack-detection
- knowledge-distillation
model-index:
- name: YOLO26s_RDD_FRDC_Distilled_v2
results:
- task: {type: object-detection, name: Road Damage Detection}
dataset: {name: Unified Road Defect Dataset (held-out val, 4509 imgs), type: TamAko783/Unified_Road_Defect_Dataset}
metrics:
- {type: mAP50, value: 0.692, name: mAP@50}
- {type: mAP50-95, value: 0.375, name: mAP@50-95}
- {type: F1, value: 0.672, name: F1}
---
# YOLO26s · Distilled v2 (small, 2 teachers)
A **YOLO26s** road-damage detector (4-class CRDDC: D00 longitudinal, D10 transverse,
D20 alligator, D40 pothole) on the [Unified Road Defect Dataset](https://huggingface.co/datasets/TamAko783/Unified_Road_Defect_Dataset).
**Method:** **Semi-supervised distillation, two teachers** (Co-DETR + RTMDet, WBF) on the larger YOLO26s student. Best overall model in the suite.
## Metrics — RDD held-out validation (4,509 images, 11,470 boxes)
| Class | mAP@50 | mAP@50-95 | Precision | Recall | F1 |
|---|---:|---:|---:|---:|---:|
| D00 Longitudinal | 0.635 | 0.357 | 0.720 | 0.561 | 0.630 |
| D10 Transverse | 0.664 | 0.345 | 0.715 | 0.589 | 0.646 |
| D20 Alligator | 0.727 | 0.409 | 0.743 | 0.647 | 0.692 |
| D40 Pothole | 0.741 | 0.390 | 0.774 | 0.668 | 0.717 |
| **Overall** | **0.692** | **0.375** | **0.738** | **0.616** | **0.672** |
## Full model comparison (same held-out val)
All five models in this study, evaluated identically (imgsz 640):
| Model | Variant | Params | Distillation | mAP@50 | mAP@50-95 | F1 |
|---|---|---:|---|---:|---:|---:|
| [YOLO26n_RDD_Base](https://huggingface.co/TamAko783/YOLO26n_RDD_Base) | YOLO26n | 2.4M | — (GT only) | 0.635 | 0.334 | 0.621 |
| [YOLO26n_RDD_FRDC_Distilled](https://huggingface.co/TamAko783/YOLO26n_RDD_FRDC_Distilled) | YOLO26n | 2.4M | 1 teacher (Co-DETR) | 0.640 | 0.337 | 0.625 |
| [YOLO26n_RDD_FRDC_Distilled_v2](https://huggingface.co/TamAko783/YOLO26n_RDD_FRDC_Distilled_v2) | YOLO26n | 2.4M | 2 teachers (Co-DETR+RTMDet) | 0.638 | 0.337 | 0.626 |
| [YOLO26s_RDD_Base](https://huggingface.co/TamAko783/YOLO26s_RDD_Base) | YOLO26s | 9M | — (GT only) | 0.687 | 0.372 | 0.665 |
| **➤ YOLO26s_RDD_FRDC_Distilled_v2** (this model) | YOLO26s | 9M | 2 teachers (Co-DETR+RTMDet) | **0.692** | **0.375** | **0.672** |
**Reading it:**
- **Distillation helps** — every distilled model beats its GT-only baseline.
- **Capacity helps most** — the YOLO26s models (+~0.05 mAP@50) clearly outperform the nano ones on this val.
- **One vs two teachers** is a near-tie at nano size; the two-teacher set's edge is small.
- **Cross-domain check (independent RDDC2024-ID, Indonesia, 8,901 imgs):** distilled models
generalized *better* than baselines, while the larger GT-only model generalized *worse* —
evidence the distillation's added data improves robustness, not just in-domain fit.
> RDD ground truth has known missing annotations, so absolute precision/recall are conservative
> for all models. The comparison is fair — every model uses the identical held-out val, never trained on.
## Usage
```python
from ultralytics import YOLO
model = YOLO("YOLO26s_RDD_FRDC_Distilled_v2.pt")
results = model("road.jpg")
```
## Models in this suite
- [YOLO26n_RDD_Base](https://huggingface.co/TamAko783/YOLO26n_RDD_Base) · [YOLO26n_RDD_FRDC_Distilled (v1)](https://huggingface.co/TamAko783/YOLO26n_RDD_FRDC_Distilled) · [YOLO26n_RDD_FRDC_Distilled_v2](https://huggingface.co/TamAko783/YOLO26n_RDD_FRDC_Distilled_v2)
- [YOLO26s_RDD_Base](https://huggingface.co/TamAko783/YOLO26s_RDD_Base) · [YOLO26s_RDD_FRDC_Distilled_v2](https://huggingface.co/TamAko783/YOLO26s_RDD_FRDC_Distilled_v2)
## Datasets
- Base: [TamAko783/Unified_Road_Defect_Dataset](https://huggingface.co/datasets/TamAko783/Unified_Road_Defect_Dataset)
- Distillation sets: [v1](https://huggingface.co/datasets/TamAko783/Unified_Road_Defect_FRDC_Pseudolabeled) ·
[v2 (two-teacher)](https://huggingface.co/datasets/TamAko783/Unified_Road_Defect_FRDC_Pseudolabeled_v2)
## Credits
- Datasets: RDD-2022 (Arya et al.) · UAV-PDD2023 · RoadDamageVision (Silva Zendron & Leithardt, CC BY 4.0).
- Distillation teachers: **Co-DETR Swin-L + RTMDet-x** — FRDC (Wang Fangjun et al.), ORDDC'2024 winner.
- Student: YOLO26s, AGPL-3.0 (Ultralytics).
|