i.
Generative 3D & Novel Views
Lifting 2D diffusion foundation models into 3D — wavelet-based geometry priors and multi-view consistency for photorealistic scene synthesis.
§ 01
about
Hello — I'm Xueyang Kang,
a researcher at the
intersection of 3D vision
& embodied robotics.
I am a Research Fellow at Nanyang Technological University (School of EEE), working on robotic perception and 3D scene understanding for embodied AI — bringing together graph neural networks, equivariant geometry learning, and generative 3D models.
I completed a joint Ph.D. at the University of Melbourne and KU Leuven (2021–2025), preceded by Master's degrees from TU Munich and Tongji University. Before academia, I built production perception systems at Qualcomm and Momenta AI, with research stints at Telstra IoT Lab and HK PolyU.
News
-
Joined Nanyang Technological University as a Research Fellow in Electrical & Electronics Engineering.
-
Paper accepted at CVPR 2026 — Hierarchical Point-Patch Fusion with Adaptive Patch Codebook for Shape Anomaly Detection.
-
Ph.D. defended at the University of Melbourne. Thesis: Geometric Deep Learning.
-
Completed a six-month research internship at Telstra IoT Smart Sensor Research Lab (Melbourne) — radar-based infrastructure monitoring.
-
Paper accepted at ACM MM 2025 as ORAL — Look Beyond: Two-Stage Scene View Generation via Panorama and Video Diffusion.
-
Joint-Ph.D. completed at KU Leuven.
-
Paper accepted at ECCV 2024 as ORAL — Equi-GSPR: top <2% of 8,300 submissions.
-
FocDepthFormer accepted as Oral at AJCAI 2024.
-
Began a four-month research visit at The Hong Kong Polytechnic University (AAE Faculty).
-
Adaptive Sampling-based Particle Filter for Visual-inertial Gimbal in the Wild accepted at ICRA 2023.
Education
-
Joint Ph.D. · University of Melbourne & KU Leuven.
Engineering & Information Technology · Electrical & Information Engineering -
M.Sc. · Technical University of Munich (TUM).
Electrical & Information Engineering · GPA 1.7/1.0 -
M.Sc. · Tongji University, Shanghai.
Electrical & Information Engineering · GPA 86.5/100
§ 02
research
Teaching robots to see, reconstruct, & imagine 3D space.
My research builds robust, generalizable 3D perception for embodied agents — from equivariant geometry learning that survives rotation and scale, to generative models that complete and synthesize scenes from sparse evidence.
ii.
Reconstruction & Registration
SE(3)-equivariant graph networks for sparse point-cloud registration; focal-stack and RGB-D depth that generalises across modalities.
iii.
Geometry Representation
Implicit SDFs, manifold graph representations, and Gaussian splats unified for completion, anomaly detection, and interactive 3D segmentation.
iv.
Robotic Fusion Sensing
Multi-modal fusion of cameras, IMU, LiDAR, UWB, radar & wheel encoders — VIO, EKF state estimation, and embedded deployment.
Highlights
▦ point-patch fusion
Hierarchical Point-Patch Fusion
Adaptive patch codebook for 3D shape anomaly detection.
CVPR 2026 →▦ panorama diffusion
Look Beyond
Two-stage scene view generation via panorama + video diffusion.
ACM MM 2025 · Oral →▦ equi-gspr
Equi-GSPR
SE(3)-equivariant graph net for sparse point-cloud registration.
ECCV 2024 · Oral · Top <2% →▦ particle filter VIO
Adaptive Particle Filter VIO
Visual-inertial gimbal estimation in unstructured outdoor scenes.
ICRA 2023 →
§ 03
publications
Selected papers.
★ first or corresponding author. Full list on Google Scholar & ORCID.
▦
Hierarchical Point-Patch Fusion with Adaptive Patch Codebook for Shape Anomaly Detection
CVPR 2026 CORE A*
▦
Look Beyond: Two-Stage Scene View Generation via Panorama and Video Diffusion
ACM MM 2025 CORE A* · Oral paper
▦
▦
Multi-view Geometry-Aware Diffusion Transformer for Indoor Novel View Synthesis
ICLR 2025 Workshop IJCNN 2025 Oral paper
▦
Surfel-based 3D Registration with Equivariant Features
CVPR Equi-Vision Workshop 2025 IGARSS 2025 poster
▦
FocDepthFormer: Transformer with Latent LSTM for Depth Estimation from Focal Stack
AJCAI 2024 Oral paper
▦
▦
3D Reconstruction & Assessment Framework based on Affordable 2D LiDAR
IEEE/ASME AIM 2018 Oral IEEE link
Under Review
-
Zero Shot Style Transfer to Gaussian Splatting.
-
Wavelet-based Geometry Prior from 2D Diffusion Foundation Model for High-Quality 3D Reconstruction (supervised by Prof. Guo Yulan).
-
Very Few Click-based Interactive 3D Segmentation with Semantic Prototype Embedding (accepted, minor revisions).
-
A Survey of Robotic Navigation & Manipulation with Physics Simulators in the Era of Embodied AI (accepted, major revisions).
-
Soft Robotic Finger for Texture Unfolding with Visual Feature Fusion (supervised by Prof. Jianwei Zhang).
-
MeshGuard: Robust and Imperceptible Watermarking of 3D Mesh Assets via Laplace–Beltrami Spectral Embedding (supervised by Prof. Daniel Cremers).
-
Robust Convex Decomposition-based Mesh Reconstruction from Point Cloud (supervised by Prof. Matthias Niessner).
§ 04
career
Industry & academia, across four continents.
Six years of industrial perception engineering — at Qualcomm, Momenta, and Telstra — woven with a joint Ph.D. between Australia and Belgium and research visits across Europe and Asia.
Positions
-
Research Fellow · Nanyang Technological University, Singapore.
School of Electrical & Electronics Engineering
Robotic perception & 3D scene understanding for embodied AI using GNNs and geometry representation learning. -
Research Intern · Telstra IoT Smart Sensor Research Lab, Melbourne.
Next-generation A121 radar sensor nodes for underground cable-well monitoring; ultra-low-power Pulsed Coherent Radar (PCR) algorithms. -
Visiting Scholar · Hong Kong PolyU (AAE Faculty).
Proposed and implemented Equi-GSPR, an SE(3)-equivariant GNN for point-cloud registration. -
Senior Algorithm Engineer · Momenta AI, Suzhou.
Autonomous-parking perception & tracking — multi-sensor fusion for obstacle avoidance, IMM filter with Ackermann kinematics, and 3D ground-line fusion. -
Robot System Engineer · Qualcomm R&D, Beijing.
VIO improved by Electronic Image Stabilization; IR+RGB feature fusion for day–night SLAM; EKF coupling of visual odometry, IMU, and wheel encoders. -
Research Assistant (HIWI) · TUM Chair of Navigation & Communication, Munich.
Integrated UWB + stereo vision in a ROS framework for drift-corrected SLAM with hardware-triggered synchronisation.
Patents
▣
Simultaneous Localization and Mapping using Cameras Capturing Multiple Spectra of Light
PCT PCT/CN2020/119769 · US20230177712A1
▣
Integrated Visual-Inertial Odometry and Image Stabilization for Image Processing
PCT PCT/CN2021/070099 · US20230421902A1
▣
Vision-based 3D Obstacle Ground-line Fusion Framework
CN CN1155123
Tools & Skills
Learning
PyTorch · PyTorch Geometric · TensorFlow · Diffusion · GNN · Transformers · Equi-models · SDF · Gaussian Splatting
Robotics & Sim
ROS · Isaac-Sim · PyBullet · Blender · Three.js · LiDAR · RGB-D · IMU · UWB · mmWave Radar
Hardware
FPGA · ARM M4 · Jetson Nano/X · Raspberry Pi · Embedded C
Languages
English · 中文 (native) · Deutsch · Nederlands (beginner)
Beyond research
Chess · films · football · table tennis · piano · hiking · drawing · travelling whenever a conference permits it.
§ 05
reach
A network spanning the globe.
Each pulse marks a place where my work has touched — from labs in Melbourne and Leuven to industry teams in Beijing, Suzhou, and Singapore, and readers across the world. Drag to rotate; the Earth turns on its own when idle.
drag to spin
§ 06
contact
Let's talk.
I'm always happy to hear about collaborations on 3D perception, geometric deep learning, or robotic sensor fusion, or to chat with prospective students, interns, and visiting researchers.
For research enquiries, please include a brief description of your project, your CV, and any prior relevant work. The fastest way to reach me is by email.
- Email kangxueyang@126.com
- Mobile +86 18221799938
- WeChat Alexandorcaisyo
- ORCID 0000-0001-7159-676X
- Affiliation School of EEE · NTU Singapore