Events

Date: 11 June 2026   Time: 15:00 - 16:00

Location: PL 4.24, Peter Landin Building

You are cordially invited to the Centre for Intelligent Sensing seminar from an external speaker:

Dr Youngkyoon Jang: Aligning Data Strategies in 3D Vision: From Virtual Humans to Neural Rendering and VLM Reasoning
When: Thursday, 11th June, 15:00-16:00
Where: PL 4.24, Peter Landin Building

Speaker:
Dr Youngkyoon Jang, (incoming) Staff ML Engineer, Rivian

Title:
Aligning Data Strategies in 3D Vision: From Virtual Humans to Neural Rendering and VLM Reasoning

Abstract:
In this talk, I will explore how aligning data strategies and representations can minimize uncertainty, bridge modalities, and enhance performance across complex computer vision and reasoning tasks. First, I will present an interactive 3D virtual human avatar (iCo3D) framework where face, body, and speech modules are aligned using a global coordinate system. This approach allows for the seamless integration of independently trained components, resulting in a photorealistic avatar capable of real-time, interactive conversations. Second, I will discuss our advancements in sparse novel view synthesis, starting from a foundational method (CoMapGS) that utilizes covisibility maps by aligning monocular depth estimation, sparse point clouds, and dense correspondences. Building upon this, I will introduce our latest iteration, SA-ResGS, which incorporates residual learning via a novel view-dependent, ray-anchored skip-connection approach to significantly improve rendering quality under active view selection scenarios. Finally, I will transition from spatial alignment to structural alignment across modalities by introducing Map2Thought. This work demonstrates how building explicit metric cognitive maps can fundamentally ground and enhance the implicit, uncertain reasoning processes of 3D Vision-Language Models (VLMs) in spatial understanding scenarios. Together, these frameworks highlight the pivotal role of structured data alignment—whether geometric, radiometric, or cognitive—in resolving ambiguity and developing robust, state-of-the-art AI solutions.

Bio:
Youngkyoon Jang is an incoming Staff ML Engineer in the Autonomy & AI team at Rivian UK in London, where he will focus on perception for 3D vision and closed-loop simulation. Previously, from November 2022 to April 2026, he was a Senior Research Scientist and Sub-project Lead in the 3D Vision Team at Huawei Technologies R&D UK (Noah's Ark Lab, London). Prior to his time at Huawei, he worked as a Software Engineer and Founding Research Specialist at Disguise, and held postdoctoral research positions at Imperial College London, the University of Bristol, Queen Mary University of London, and KAIST. He received his Ph.D. from the Korea Advanced Institute of Science and Technology (KAIST) in August 2015 under the supervision of Woontack Woo, and was co-advised by Tae-Kyun Kim at Imperial College London. His research focuses on novel visual sensing technologies designed to make interactions between humans and autonomous systems more intuitive in real-world environments. His work involves designing and curating novel datasets, modeling and analyzing visual attributes, understanding human behavior, and investigating decision-making fairness through visual computing and data analysis. Leveraging a strong foundation in computer science, his research integrates advanced machine learning and deep learning methodologies—including 3D Gaussian Splatting (3DGS), NeRF, 3D-VLMs, Transformers, and CNNs—to solve complex computer vision problems.