Semantic image segmentation

• Main Author: Melia, Selvira Junita
• Other Authors: Lu Shijian
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/171906
• Institution: Nanyang Technological University

Semantic segmentation is useful in real world applications due to its deep understanding of scene imagery, yet it is one of the most challenging problems in computer vision. The pixel-level nature of the task makes training of models expensive and time-consuming. A popular approach to overcome this issue is to use a model that was pre-trained on a different dataset. However, differences in training and test domains cause a reduction in model performance. This project addresses the problem by focusing on the task of unsupervised domain adaptation in semantic segmentation. Research into the various novel techniques within the field identified the presence of hyperparameters specific to techniques being implemented. Adjustments to such hyperparameters have the potential to provide interesting data on model performance. To investigate this research gap, this project focuses on the effects of a novel method with a category contrast technique on a selected baseline method. Experiments were conducted to measure improvements in model performance based on varying values of the weight of contrastive loss. It was concluded that a higher weight generally leads to improved model performance, although variations in scores among different classes suggest future research directions that focus on fine-tuning technique-specific hyperparameters within the unsupervised domain adaptation in semantic segmentation field.