CUBIT: A High-resolution Infrastructure Defect Dataset
Fully Evaluated with Autonomous Detection Framework
Submitted to International Conference on Acoustics, Speech, & Signal Processing 2024 (ICASSP 2024)
Benyun Zhao1, Xunkuai Zhou2, Guidong Yang1, Junjie Wen1, Jihan Zhang1, Xi Chen1, and Ben M. Chen1, IEEE Fellow
1.Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong
2.School of Electronics and Information Engineering,Tongji University
2.School of Electronics and Information Engineering,Tongji University
Learning-based visual inspection, integrated with unmanned robotic system, offers a more effective, efficient, and safer alternative for infrastructure inspection tasks that are traditionally heavily reliant on human labor. However, the potential of learning-based inspection methods remains limited due to the lack of publicly available, high-quality datasets. This paper presents CUBIT, a high-resolution defect detection dataset comprising more than 5500 images with resolutions up to 8000 x 6000 which covers a broader spectrum of practical situations, backgrounds, and defect categories than existing publicly available datasets. We conduct extensive experiments to benchmark the performance of state-of-the-art real-time detection methods on our proposed dataset, validating the effectiveness of it. Moreover, based on the benchmark results, we develop a module named GIPFPP to integrate multi-scale feature, enhancing the AP by 3% while reducing the number of parameters by 10% on baseline model. Additionally, a real-site UAV-based inspection has been conducted to verify the reliability of the dataset.
The sample images in CUBIT has been shown below. All the data are collected by autonomous unmanned systems such as UAV and UGV. Our dataset includes various scnarios and defect categories compared with the existing open-source bounding-box level defect detection dataset.
| Dataset | Num. of Images | Resolution | Data Collection Platform | Category | Scenario | Material | Experiments |
|---|---|---|---|---|---|---|---|
| RDD-2018 | 9053 | 600x600 | Smartphones | Crack, Corrosion | Pavement | Asphalt | SSD |
| RDD-2019 | 13135 | 600x600 | Smartphones | Crack, Corrosion | Pavement | Asphalt | SSD |
| RDD-2020 | 26336 | 600x600, 720x720 | Smartphones | Crack, Pothole | Pavement | Asphalt | SSD |
| RDD-2022 | 47420 | 512x512, 600x600, 720x720, 3650x2044 | Smartphones, Hand-held cameras, UAV cameras, Google street view | Crack, Pothole | Pavement | Asphalt | - |
| PID | 7237 | 640x640 | Crawled from Internet | Crack | Pavement | Asphalt | YOLOv2, Fast R-CNN |
| Murad | 2620 | up to 838x809 | Smartphones | Crack | Pavement | Asphalt | Faster R-CNN |
| CODEBRIM | 1590 | up to 6000x4000 | Hand-held cameras, UAV Cameras | Crack, Corrosion | Bridge | Concrete | MetaQNN, ENAS |
| CUBIT | 5527 | 4624x3472 and 8000x6000 | Cameras in Unmanned Systems | Crack, Spallinig, Moisture | Building (65%), Pavement (29%), Bridge (6%) | Concrete, Asphalt, Stone | Faster R-CNN, PP-YOLO, PP-YOLOv2, YOLOX, YOLOv5, YOLOv7, YOLOv6, YOLOv6+GIPFPP(ours), Real-site experiment |
The visualization of defect detection framework based on CUBIT dataset is illustrated below, which encompasses the entire process: data collection by autonomous unmanned system; the baseline network integrated with our GIPFPP module; the output of defect detection results.
We enlarge the prediction results in the bottom right corner of framework images above. CUBIT dataset covers three infrastructure types: Building, Pavement, and Bridge, and aims for three types of defect: Crack, Spalling, and Moisture. Rectangles indicate the output prediction box Red for Crack, Pink for Spalling, and Orange for Moisture with inferred defect type and confidence score from YOLOv6-n+GIPFPP trained on the training set of our proposed dataset.
Qualitative visualization of UAV-based real-world experiment is shown below. On the left, our multi-UAVs inspection schematics is illustrated. On the right, the detection results of four direction façades of the building are displayed.
