Author: Maria Vila Abad
Supervisors: Luís Herranz
Presentation time: 10:30
Virtual Room: 1.3 | Live presentation URL
Construction is one of the world’s largest but also most inefficient and wasteful industries. One of its biggest problems is that though we design digitally, we still construct manually. This causes the constructed structure to significantly diverge from the designed model. Digitally monitoring the evolution of construction sites can help closing the gap between virtual and real buildings as it allows us to quickly discover errors and deviations from the virtual model. Part of this monitoring consists on segmenting instances of building elements in images of construction sites to control their accurate matching with the virtual design. Deep Convolutional Neural Networks have been proven to succeed in this task but they need large annotated data to learn from. Manually labeling an instance segmentation dataset is costly and time-consuming and may not be viable in many occasions. In this thesis, we present two different procedures to obtain instance segmentation annotations at a reduced human supervision cost. The first approach tackles the problem by using 3D information from the virtual building model and the construction site to automatically find correspondences between them in order to obtain instance segmentation labels. The second approach consists on a Weakly Semi-Supervised method that learns to predict instance segmentation masks from image-level labels (which are cheaper to annotate) and a small number of instance segmentation masks. We obtain promising results that could lead to significant savings in annotation costs.