NCKU transforms the moment with computing to weave the future with intelligence

The real estate industry has undergone significant transformations with the emergence of innovative technologies. AI-powered solutions have played a pivotal role in enhancing the property viewing experience. This talk explores the development of an AI-Powered 3D Live Tour Application tailored specifically for the real estate industry, with a particular focus on the technology of 3D layout reconstruction from panoramas. The application harnesses the power of artificial intelligence and computer vision algorithms to create immersive and interactive virtual tours of real estate properties. By leveraging panoramic images captured through specialized cameras or smartphones, the application employs advanced 3D layout reconstruction techniques to accurately recreate the physical spaces in a virtual environment. In this presentation, I will highlight three works that tackle the challenge of 3D indoor layout reconstruction from different input settings, including single-view panorama and multi-view panoramas. I will further introduce a transformer-based network in the multi-view setting, enabling simultaneous learning of pose registration and layout reconstruction. This novel approach showcases the potential for more comprehensive and accurate property layout reconstruction, benefiting both potential buyers and real estate professionals.

為了展示學生在敏求智慧運算學院學分學程及學位學程課程的學習成果，以及院聘老師團隊在專題上的研究進展。提供一個成果展示及互動交流平台。讓學生能夠展示他們在課程或專題中所學到的知識、技能和成果，並鼓勵學生主動參與學程課程，激發他們的學習動力。

Over recent years, machine learning has revolutionized the generation of photorealistic facial images, with image-to-image transformation techniques being employed in a plethora of applications, including data augmentation, entertainment, virtual reality, and even in the synthesis of deepfake images/videos. In this talk, I plan to review the general machine learning methodologies employed for face image generation before delving into our group's novel research on cross-domain heterogeneous face generation models. I will elucidate how we model facial geometry, sketches, expressions, and lighting, and how these models enable inference about unseen test subjects. The talk will conclude with the presentation of a framework for a facial generator designed to recover faces obscured by masks, a crucial innovation for surveillance systems in the era of Covid-19.