BFD-UAV2K: Deployment-Oriented UAV Facade Defect Detection

BFD-UAV2K is a deployment-oriented benchmark for automated UAV-based facade defect detection. It contains 2,000 full-frame images acquired during coverage-oriented inspection of 42 reinforced-concrete buildings. The benchmark uses separate visual annotations for hollow, spalling, and crack, and it partitions the data by building so that no building appears in more than one subset.

The benchmark is designed for regional, multi-building screening. It connects coverage-oriented acquisition, annotation effort, class-wise detector behavior, inference efficiency, and diagnosis-guided optimization under one fixed evaluation protocol.

Dataset Overview

Statistic	Overall	Train	Validation	Test
Buildings	42	32	5	5
Images	2,000	1,600	200	200
Defect-present images	1,280	1,021	164	95
Background-only images	720	579	36	105
Defect instances	5,020	3,988	505	527

The 720 background-only images are naturally occurring facade views rather than synthetic negatives. They retain joints, stains, shadows, facade textures, and architectural components that can produce false alarms during screening.

Class Distribution

Class	Overall	Train	Validation	Test
Hollow	1,391	1,137	135	119
Spalling	2,528	1,967	256	305
Crack	1,101	884	114	103

The fixed class order is hollow, spalling, and crack in both YOLO TXT and COCO JSON annotations.

Acquisition and Annotation

Images were collected from 42 modern reinforced-concrete buildings with DJI Mini 3 and DJI Matrice 4T UAVs. Pilots followed predefined facade-coverage sequences and used a hover-capture-move procedure to obtain full-frame, facade-facing images. The collection includes marble, cement, and concrete finishes under realistic field variation.

Three annotators with civil-engineering knowledge performed the initial labeling in LabelImg, and two reviewers checked class assignment, box extent, missing instances, and ambiguous cases. The three visual categories are defined as follows:

• Hollow: visible local bulging, unevenness, lifting, or separation of the finish layer. This label is a visual screening category and does not confirm a subsurface cavity, depth, cause, or severity.
• Spalling: visible loss or detachment of surface material.
• Crack: visually identifiable linear fractures.

Only defects visible with sufficient confidence were annotated. Ambiguous regions were excluded instead of being assigned speculative labels.

Building-Disjoint Split

The split is fixed at 32/5/5 buildings for training, validation, and testing. Because consecutive coverage images can share wall regions, lighting, textures, architectural components, or repeated views of the same defect, an image-level random split could leak highly correlated observations across subsets. Building-level partitioning instead evaluates detectors on buildings excluded from model development.

Benchmark Protocol

Eight detectors from three architecture families were trained from COCO-pretrained initialization for 200 epochs at an input size of 1024 x 1024. Validation AP50:95 selected the checkpoint, while the held-out test buildings were reserved for final evaluation.

Family	Models
One-stage	YOLOv5mu, YOLO11n, YOLO11m, YOLO26m
Transformer	RT-DETR-L, RT-DETR-X
Two-stage	Faster R-CNN, Cascade R-CNN

The benchmark also evaluates five labeled-image budgets: 100, 200, 400, 800, and 1,600 training images. The 100-800 settings use five nested subset seeds; the 1,600-image setting is the full-data reference.

Full-Data Results

The table reports class-wise AP50 on the held-out test buildings and standardized single-GPU efficiency. Timing used one NVIDIA GeForce RTX 4090 at 1024 x 1024, batch size 1, and FP32.

Model	Hollow AP50	Spalling AP50	Crack AP50	Mean ms/image	FPS
YOLOv5mu	0.372	0.728	0.195	7.58	131.91
YOLO11n	0.363	0.638	0.187	5.90	169.62
YOLO11m	0.405	0.730	0.161	7.97	125.47
YOLO26m	0.429	0.722	0.202	8.26	121.11
RT-DETR-L	0.316	0.665	0.235	38.11	26.24
RT-DETR-X	0.256	0.735	0.221	43.32	23.08
Faster R-CNN	0.258	0.594	0.201	27.55	36.30
Cascade R-CNN	0.304	0.611	0.185	31.65	31.60

Spalling is the most detectable class across model families, whereas crack remains the most difficult. The results do not support one universal model ranking: YOLO11n is the lightweight throughput reference, YOLO26m has the highest hollow AP50, RT-DETR-X has the highest spalling AP50 and recall, and RT-DETR-L has the highest crack AP50.

Annotation Scale and Effort

Class-wise learning curves show that the value of additional labels depends on the defect category and detector family. In particular, crack performance remains model- and seed-dependent across the evaluated budgets, so the paper does not propose a universal minimum dataset size.

Annotation and review were recorded for the first 500 images. Initial labeling required 6.23 person-hours and review required 4.63 person-hours, for a combined 10.87 person-hours. Values for the five benchmark budgets are linear planning estimates beyond the recorded 500 images.

Images	Initial labeling (h)	Review (h)	Combined (h)
100	1.25	0.93	2.17
200	2.49	1.85	4.35
400	4.99	3.71	8.69
800	9.97	7.41	17.39
1,600	19.95	14.83	34.77

Diagnosis-Guided Optimization

Validation diagnosis identifies recurring difficulty in small targets, elongated targets, defect-dense scenes, mixed-class scenes, and background-only images. Multi-scale training, hard-negative mining, geometry-aware cropping, and a combined multi-scale plus hard-negative variant were then evaluated with YOLO26m, RT-DETR-L, and Cascade R-CNN.

The effects are architecture- and class-dependent. Consequently, these methods should be treated as diagnosis-matched candidates rather than general upgrades.

Download

The release is provided as a multipart archive:

BFD-UAV2K_public_release.part1.rar
BFD-UAV2K_public_release.part2.rar
BFD-UAV2K_public_release.part3.rar

Download all three parts into the same directory and begin extraction from BFD-UAV2K_public_release.part1.rar.

Release Checksums

File	SHA256
BFD-UAV2K_public_release.part1.rar	305B87DD9551EAABFD2327129A97C1C35C082B54E60DFA4CB7309F9FC5F90B75
BFD-UAV2K_public_release.part2.rar	23FF814A62C2FF412A2B269D24BC824EA9F782514A7FB87AEF47C78F2CE837BC
BFD-UAV2K_public_release.part3.rar	C7A2730F34E017870140BD7AB3C17B7C1DCB9D098F6DAED42F1000BD787682F3

Paper

Yu Xia, Boyang Zhang, Zhihong Pan, Kang Gao, Hanbo Yang, Ruoyu Chen, and Kang Yang. "Deployment-oriented benchmark for automated UAV-based facade defect detection." Automation in Construction, 193 (2027), 107280. https://doi.org/10.1016/j.autcon.2026.107280

@article{xia2027deployment,
  title   = {Deployment-oriented benchmark for automated UAV-based facade defect detection},
  author  = {Xia, Yu and Zhang, Boyang and Pan, Zhihong and Gao, Kang and Yang, Hanbo and Chen, Ruoyu and Yang, Kang},
  journal = {Automation in Construction},
  volume  = {193},
  pages   = {107280},
  year    = {2027},
  doi     = {10.1016/j.autcon.2026.107280}
}

License

UAV2K is released under the UAV2K Academic Research License.

The dataset is available for non-commercial academic research, scientific research, and educational purposes only.

Users may preprocess, transform, partition, and augment the dataset for internal research and experimentation. However, redistribution, re-hosting, or republication of the original dataset, modified versions, or dataset subsets is not permitted without prior written permission from the authors.

Commercial use requires separate authorization from the UAV2K authors.

Users of UAV2K are required to appropriately cite the corresponding publication and should obtain the dataset directly from the official repository.

For the complete license terms, please see the LICENSE file.

Contact

• Ruoyu Chen: chenruoyu@just.edu.cn
• Kang Yang: yangkg001@outlook.com