Data Augmentation
Dr. Yuhan Jiang 2020/11/04
Two approaches to augment the datasets for training deep learning models [using few orthographic (front, back, left and right) view images]:
Applying image perspective transformation (randomly) to augment the training datasets (orthographic view images which are collected from a 3D mesh model or point cloud, and their manually crafted labels). With this process, the trained model can work with arbitrary views of the 3D mesh model / point cloud.
See the following examples
Left Side fixed, while the right side changed (with image width changed)
Upper Side fixed, while the bottom side changed (with image height changed)
2. Applying image brightness / color/ contrast/ sharpness adjustment
See the following examples
Reference:
Jiang, Y., Han, S., and Bai, Y. (2022). “Scan4Façade: Automated As-Is Façade Modeling of Historic High-Rise Buildings Using Drones and AI.” ASCE Journal of Architectural Engineering, 28(4), DOI: 10.1061/(ASCE)AE.1943-5568.0000564.
Jiang, Y., Han, S., Li, D., Bai, Y., and Wang, M. (2022). “Automatic Concrete Sidewalk Deficiency Detection and Mapping with Deep Learning.” Expert Systems with Applications, 207 (117980), DOI: 10.1016/j.eswa.2022.117980.
Jiang, Y., Huang, Y., Liu, J., Li, D., Li, S., Nie, W., Chung, I*. (2022). “Automatic Volume Calculation and Mapping of Construction and Demolition Debris Using Drones, Deep Learning, and GIS.” Drones, 6(10), 279, DOI: 10.3390/drones6100279
Paul Haeberli and Douglas Voorhies, "Image Processing By Interpolation and Extrapolation", http://www.graficaobscura.com/interp/index.html