EngJamesO commited on
Commit
cf8fc78
·
verified ·
1 Parent(s): feda0df

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +87 -1
README.md CHANGED
@@ -12,4 +12,90 @@ tags:
12
  library_name: ultralytics
13
  language:
14
  - en
15
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
  library_name: ultralytics
13
  language:
14
  - en
15
+ ---
16
+
17
+ ---
18
+ license: mit
19
+ tags:
20
+ - object-detection
21
+ - yolo
22
+ - computer-vision
23
+ - pothole-detection
24
+ - road-infrastructure
25
+ - ultralytics
26
+ library_name: ultralytics
27
+ ---
28
+
29
+ # Pothole Detection Models
30
+
31
+ This repository contains trained YOLO models for detecting potholes in road images. The models were developed to assist with road infrastructure monitoring and maintenance planning.
32
+
33
+ ## Model Overview
34
+
35
+ Five YOLO architectures were trained and evaluated for pothole detection:
36
+
37
+ | Model | Parameters | mAP50 | mAP50-95 | Latency (ms) | FPS |
38
+ |-------|-----------|-------|----------|--------------|-----|
39
+ | YOLOv9c | 25.3M | 0.792 | 0.485 | 36.3 | 27.6 |
40
+ | YOLOv12n | 2.56M | 0.785 | 0.463 | 13.9 | 72.1 |
41
+ | YOLOv8n | 3.01M | 0.782 | 0.465 | 6.9 | 144.7 |
42
+ | YOLOv11 | 2.58M | 0.779 | 0.460 | 9.3 | 107.0 |
43
+ | YOLOv8-FPN | 2.21M | 0.707 | 0.396 | 6.4 | 155.8 |
44
+
45
+ ## Intended Uses
46
+
47
+ - **Primary Use**: Automated pothole detection in road images for infrastructure assessment
48
+ - **Secondary Use**: Research in object detection, computer vision applications for transportation
49
+ - **Target Users**: Road maintenance departments, urban planners, researchers
50
+
51
+ ## Model Selection Guide
52
+
53
+ - **YOLOv9c**: Best overall accuracy (mAP50: 0.792), recommended when accuracy is prioritized over speed
54
+ - **YOLOv8n**: Good balance of accuracy and speed, suitable for real-time applications
55
+ - **YOLOv12n**: Lightweight with competitive accuracy, ideal for edge deployment
56
+ - **YOLOv11**: Efficient architecture with modern optimizations
57
+ - **YOLOv8-FPN**: Fastest inference, suitable for high-throughput scenarios
58
+
59
+ ## Performance at Different Confidence Thresholds
60
+
61
+ Based on qualitative evaluation (40 test images, 86 ground truth objects):
62
+
63
+ ### YOLOv9c (Recommended)
64
+ - **Conf=0.25**: Precision: 0.804, Recall: 0.831, F1: 0.790
65
+ - **Conf=0.50**: Precision: 0.846, Recall: 0.723, F1: 0.753
66
+ - **Conf=0.70**: Precision: 0.700, Recall: 0.505, F1: 0.558
67
+
68
+ ### YOLOv8n
69
+ - **Conf=0.25**: Precision: 0.807, Recall: 0.838, F1: 0.791
70
+ - **Conf=0.50**: Precision: 0.840, Recall: 0.710, F1: 0.751
71
+ - **Conf=0.70**: Precision: 0.600, Recall: 0.437, F1: 0.482
72
+
73
+ ## Usage
74
+
75
+ ### Installation
76
+
77
+ ```bash
78
+ pip install ultralytics opencv-python
79
+ ```
80
+
81
+ ### Python inference
82
+
83
+ ```python
84
+ from ultralytics import YOLO
85
+
86
+ # Load model
87
+ model = YOLO("yolov9c.pt")
88
+
89
+ # Run inference
90
+ results = model("path/to/image.jpg")
91
+
92
+ # Visualize results
93
+ results[0].show()
94
+ ```
95
+
96
+ ### Limitations
97
+
98
+ - Models trained on specific road conditions; performance may vary on different terrain or weather conditions
99
+ - Detection accuracy depends on image quality and resolution
100
+ - Models may have difficulty with very small or heavily obscured potholes
101
+ - Performance not validated on night-time or extreme weather images