Seeing the Forest Behind the Trees: A Novel Method for Generating Data for Overlapping Object Segmentation
Name
Karl Suurkaev
Abstract
Computer vision is a rapidly developing academic field which is gaining traction throughout different fields of expertise. Deep learning and artificial neural networks are at the forefront of recent developments but many problems remain unsolved. One of the prominent problems is the detection and segmentation of overlapping objects from images. This thesis proposes a novel layered data acquisition approach for images with overlapping objects which aims to improve the ground truth data quality. The data was generated using a custom built robotic system. The resulting dataset was tested against the U-Net and YOLOv5 artificial neural networks. Additionally, the same network models were trained on a directly annotated dataset for better results comparison. The thesis also investigated if this new data acquisition approach could be automated using artificial neural networks. The results showed that the novel approach is on par with the direct approach but allows automation of ground truth data generation. This potentially allows easy generation of large datasets which improves model quality through substantially larger quantities of training data.
Graduation Thesis language
English
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Tõnis Laasfeld, Kaspar Hollo, Dmytro Fishman
Defence year
2023