Towards high-quality panoptic segmentation

• Main Author: Chen, Chongsong
• Other Authors: Chen Change Loy
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/138017
• Institution: Nanyang Technological University

Panoptic segmentation is a recently proposed task that unifies both instance and semantic segmentation. It provides a holistic solution to scene parsing by predicting instance labels and pixel-level classification. To improve the performance of our panoptic segmentation system, we explore various methods which will be described in later part of this report. We demonstrate in our report that the understanding of instance occlusion, the joint improvement by hybrid-task learning, and the study of panoptic segmentation metrics all play crucial roles. We also participated in Joint COCO and Mapillary Workshop at ICCV 2019. On test-dev dataset split, our ensemble model achieved PQ=53.5, ranked the 1st place (without external dataset) and the 2nd place (overall).