File size: 4,953 Bytes
67b32cb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
---
pretty_name: BFD-UAV2K
task_categories:
- object-detection
language:
- en
tags:
- uav
- building-facade
- defect-detection
- object-detection
- computer-vision
size_categories:
- 1K<n<10K
---

# BFD-UAV2K: UAV-Based Building Facade Defect Detection

**BFD-UAV2K** is a realistic UAV-based building facade defect dataset and benchmark for detecting facade defects in full-frame inspection images. The dataset targets practical UAV inspection settings where defects are often small, sparse, low-contrast, and embedded in complex facade backgrounds.

<p align="center">
  <img src="assets/pipeline.png" alt="BFD-UAV2K pipeline and benchmark framework" width="92%">
</p>

## Highlights

- **2,000 full-frame UAV images** collected from real-world building facade inspection campaigns.
- **Fixed benchmark split:** 1,600 training images, 200 validation images, and 200 test images.
- **Single detection class:** `defect`, covering cracking, spalling, and hollow-drum-like facade damage.
- **Realistic inspection scenes:** marble panels, windows, concrete/cement surfaces, pipes, cables, air-conditioning units, and other facade-side interference.
- **Multiple detector families:** YOLO, RT-DETR, Faster R-CNN, and Cascade R-CNN are evaluated under a unified benchmark protocol.
- **Deployment-oriented analysis:** speed, localization quality, low-data behavior, and hard-case failure patterns are analyzed together.

## Dataset Overview

| Item | Value |
|---|---:|
| Images | 2,000 |
| Positive / negative images | 912 / 1,088 |
| Defect instances | 2,664 boxes |
| Split | 1,600 / 200 / 200 |
| Annotation format | YOLO TXT + COCO JSON |
| Class | `defect` |
| UAV platform | DJI Mini 3 |
| Capture range | 2-30 m wall-facing UAV inspection geometry |

<p align="center">
  <img src="assets/positive_negative_examples.png" alt="Positive and negative examples" width="82%">
</p>

## Benchmark Models

The benchmark compares representative detectors from three major detection families:

| Family | Models |
|---|---|
| One-stage detectors | YOLOv5mu, YOLOv8m, YOLOv11m |
| Transformer detector | RT-DETR-L |
| Two-stage detectors | Faster R-CNN, Cascade R-CNN |

## Main Results

| Model | mAP@0.5 | mAP@0.5:0.95 | Precision | Recall | F1 | Time (ms/img) |
|---|---:|---:|---:|---:|---:|---:|
| YOLOv5mu | **0.7193** | 0.3631 | 0.8043 | 0.6491 | 0.7184 | **6.00** |
| YOLOv8m | 0.6710 | 0.3347 | 0.7753 | 0.5965 | 0.6743 | 6.65 |
| YOLOv11m | 0.6799 | 0.3416 | **0.8059** | 0.5827 | 0.6763 | 7.00 |
| RT-DETR-L | 0.7116 | **0.3811** | 0.7781 | **0.6772** | **0.7241** | 14.84 |
| Faster R-CNN | 0.5750 | 0.2810 | 0.5752 | 0.6526 | 0.6115 | 12.41 |
| Cascade R-CNN | 0.5740 | 0.2960 | 0.5735 | 0.6491 | 0.6090 | 25.77 |

YOLOv5mu provides the strongest accuracy-speed trade-off, while RT-DETR-L achieves the best strict-IoU localization score and F1 score. Two-stage detectors are more sensitive to repetitive facade textures and weak defect boundaries in this benchmark.

<p align="center">
  <img src="assets/qualitative_comparison.png" alt="Qualitative comparison across baseline detectors" width="96%">
</p>

## Low-Data Behavior

The benchmark also evaluates model behavior under reduced training sizes: 100, 200, 400, and 800 samples. YOLO-family detectors show stable improvement as more data becomes available, while RT-DETR-L is more sensitive in very low-data regimes but becomes highly competitive with more labeled samples.

<p align="center">
  <img src="assets/data_scale_curves.png" alt="Detector performance under different training data sizes" width="90%">
</p>

## Hard Cases

BFD-UAV2K includes practical facade inspection challenges that expose clear differences between detector families:

| Hard-case pattern | Ratio |
|---|---:|
| Long cracks with extreme aspect ratio | 10% |
| Defect-background texture overlap | 20% |
| Multi-anomaly scenes with missed detections | 30% |
| Uneven concrete coating / non-uniform background | 30% |
| Complex structural background interference | 10% |

<p align="center">
  <img src="assets/failure_cases.png" alt="Failure case analysis" width="96%">
</p>

## Repository Contents

```text
.
|-- README.md
|-- assets/
|   |-- pipeline.png
|   |-- positive_negative_examples.png
|   |-- qualitative_comparison.png
|   |-- data_scale_curves.png
|   `-- failure_cases.png
|-- data/
|-- annotations/
|-- labels/
`-- scripts/
```

The public dataset release is hosted in this Hugging Face dataset repository. The companion GitHub repository is available at [HanboYangC/BFD-UAV2K](https://github.com/HanboYangC/BFD-UAV2K).

## Citation

```bibtex
@article{bfduav2k,
  title   = {BFD-UAV2K: A Realistic Dataset and Benchmark for Unmanned Aerial Vehicle-Assisted Building Facade Defect Detection},
  author  = {},
  journal = {},
  year    = {}
}
```

## License

License information will be added with the official public release.

## Organizer

Kang Yang, email: kyang16866@gmail.com
Ruoyu Chen, email: