الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 107 |  |  |
| | | from | 107 | | |
Abstract
This thesis explores the complex world of panoptic segmentation, a multifaceted approach that combines semantic and instance segmentation. It offers a detailed understanding of multifarious scenes, which is integral to numerous applications. However, real-time implementation of this sophisticated technique is encumbered by challenges due to the computational intensity of the involved deep learning models. The thesis focuses on the examination and refinement of the ”You Only Segment Once” (YOSO) architecture, which has made significant waves in the space of panoptic segmentation. A novel model, Real-Time YOSO (RT-YOSO), is introduced as an enhanced version, meticulously crafted for real-time panoptic segmentation tasks. The innovation is epitomized by its seamless integration with the Short-Term Dense Concatenation Networks (STDC) architecture, which enhances computational efficiency and expedites real-time performance while maintaining segmentation accuracy.
The thesis also presents an in-depth analysis of the STDC architecture, which is instrumental in augmenting the real-time performance of panoptic segmentation models. This exploration enriches the contextual framework and offers readers a panoramic view of the evolutionary journey and contributions of these diverse models to the enriched landscape of computer vision.
Experiments on the renowned Cityscapes dataset demonstrate the robustness and precision of the RT-YOSO model, attaining impressive Panoptic Quality (PQ) metrics of 59.2% and 51.2%, and a heightened frame rate of 18.5 and 27.8 FPS, respectively. This achievement marks a significant milestone in mitigating the persistent accuracy-speed dilemma in real-time applications.
In conclusion, this thesis offers architectural enhancements, detailed model analysis, and empirical validation to tackle real-time challenges in panoptic segmentation. The unveiled RT-YOSO model and its examination provide both theoretical and practical advancements, influencing future research and applications demanding rapid and accurate image segmentation.