| Popis: |
The detection of windows and their states from building façade images is crucial for applications such as digital twins and building performance simulations. Deep learning using object detection algorithms can achieve this with reliable accuracy, but its success depends on diverse and accurate datasets. While image augmentation can enhance dataset diversity, its impact on detecting windows and their states from building façade images remains understudied. This research introduces nine augmentation techniques: brightness, hue, saturation, contrast, scale, translation, cutout, weather, and combinations thereof. These techniques are evaluated using Faster R-CNN, YOLOv8, and R-FCN across various architectures, including ResNet-101, ResNet-152, MobileNetV3, EfficientNetV2, and InceptionV3, with AP50 and IoU as performance metrics. Results show that the most effective augmentation techniques vary by architecture and performance metric, highlighting the need for tailored augmentation strategies. Combining multiple techniques does not always outperform using individual methods, suggesting that augmentation strategies should be specifically adapted to each architecture. These findings indicate that the proposed augmentation techniques can improve window and window state detection and may also extend to related tasks such as window type detection and building age classification. |
