Semantic image segmentation

• Main Author: Manikkath, Bharat
• Other Authors: Lu Shijian
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Semantic segmentation; Image segmentation
• Online Access: https://hdl.handle.net/10356/175376
• Institution: Nanyang Technological University

Semantic Image Segmentation is a Computer Vision task aimed at creating pixel-level labels in images for a detailed in-depth scene understanding and therefore has many practical applications in tasks like scene understanding for self-driving cars, virtual image searches, satellite image classification and many more. Most models are trained to only perform on a limited domain and may struggle when scaled to recognize a large number of distinct classes. To deal with this problem, the zero-shot methodology aims to generalize information learned from seen classes to recognize and segment previously unseen classes without supervision. The Recent years have seen a remarkable improvement in the quality and generalizability of Vision Foundational Models. This project aims to investigate the ability of such Vision Foundational Models in performing the task of semantic segmentation without additional fine-tuning.