R3: 3D Reconstruction via Relative Regression
arXiv, 2026
project page / arXiv / code
Assembles confidence-guided pairwise poses into trajectories, enabling a 0.3B model to match 1B baselines.
Xingyu Chen
陈星宇
I am a PhD student jointly enrolled at Westlake University and Zhejiang University, in the Inception3D Lab, co-supervised by Anpei Chen and Andreas Geiger.
I am fortunate to work with Yuliang Xiu at Endless AI Lab and interned at Tencent AI Lab, collaborating with Xuan Wang and Qi Zhang. I got my M.S. from Xi'an Jiaotong University and my B.S. from Chongqing University.
I work on spatial intelligence across computer vision, machine learning, computer graphics, and robotics.
Research
The world we see is constantly changing: how do intelligent systems generalize to new observations? This question led me to quest for an understanding of the mechanisms underlying spatial intelligence and to develop methods for enabling artificial intelligence with this remarkable capability.
Specifically, I am investigating how generalizability can emerge from reusable 3D & 4D representations, how these representations of the dynamic 3D world could be learned from images & videos, and how inductive biases could serve as expert knowledge to reduce unknown parameters and make learning more efficient.
Equal Contribution *, Corresponding Author †, Project Lead ⚑
GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
European Conference on Computer Vision (ECCV), 2026
A global scene tokenizer that moves beyond pixel-wise redundant tokens.
Motion 3-to-4: 3D Motion Reconstruction for 4D Synthesis
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
project page / arXiv / code
Synthesising 4D dynamic objects from single monocular video.
TTT3R: 3D Reconstruction as Test-Time Training
TTT3R: 3D Reconstruction as Test-Time Training
International Conference on Learning Representations (ICLR), 2026
China3DV 2026 Top 5 Paper
project page / arXiv / code
A simple state update rule to enhance length generalization for CUT3R.
Human3R: Everyone Everywhere All at Once
Human3R: Everyone Everywhere All at Once
International Conference on Learning Representations (ICLR), 2026
project page / arXiv / code / interactive demo
Online human-scene reconstruction in One model, One stage.
🦣Easi3R: Estimating Disentangled Motion from DUSt3R Without Training
IEEE/CVF International Conference on Computer Vision (ICCV), 2025
project page / arXiv / code / interactive demo
Disentangles DUSt3R attention maps and repurposes them for training-free 4D reconstruction.
Feat2GS: Probing Visual Foundation Models with Gaussian Splatting
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
project page / arXiv / video / code / demo / gallery
Uses novel-view synthesis to probe texture and geometry awareness in visual foundation models.
L2G-NeRF: Local-to-Global Registration for Bundle-Adjusting Neural Radiance Fields
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023
project page / arXiv / paper / code / supplementary / video / poster
Couples local and global alignment with differentiable solvers for robust bundle-adjusting NeRF.
UV Volumes for Real-time Rendering of Editable Free-view Human Performance
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023
project page / arXiv / paper / code / supplementary / video / poster
Separates high-frequency human appearance from 3D volume and encodes it as 2D textures for real-time rendering and retexturing.
PROCA: Place Recognition under Occlusion and Changing Appearance via Disentangled Representations
IEEE International Conference on Robotics and Automation (ICRA), 2023
arXiv / paper / code / video / poster
Disentangles place, appearance, and occlusion factors, then uses the place code as a retrieval descriptor.
Sparse Semantic Map-Based Monocular Localization in Traffic Scenes Using Learned 2D-3D Point-Line Correspondences
IEEE Robotics and Automation Letters (RA-L), 2022
Estimates camera poses from sparse semantic maps through learned 2D-3D point-line correspondences.
Ha-NeRF😆: Hallucinated Neural Radiance Fields in the Wild
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022
project page / arXiv / paper / supplementary / code / video / poster
Recovers NeRF from tourist photos with variable appearance and occlusions, enabling occlusion-free renderings with hallucinated appearance.
Using Detection, Tracking and Prediction in Visual SLAM to Achieve Real-time Semantic Mapping of Dynamic Scenarios
IEEE Intelligent Vehicles Symposium (IV), 2020
Detects only keyframes and predicts dynamic objects in remaining frames for efficient semantic mapping.
Navigation Command Matching for Vision-based Autonomous Driving
IEEE International Conference on Robotics and Automation (ICRA), 2020
Matches navigation commands with smooth rewards to discriminate sub-optimal driving actions.
Projects
I am passionate about bridging the physical and digital worlds by building next-generation AR and robotics.
Kuafu (Autonomous Driving)
- GPS-Denied Navigation
- Intelligent Vehicle Challenge
Robotic Hand
- Hand gesture recognition
- Multi-sensor fusion
- Robotic hand controller
Robotic Arm
- Teleoperation
- Hand gesture recognition
- Four-bar linkage structure
Talks
Inferring the physical world and camera poses from images
Sharing the intuition of dealing with dynamic objects in our previous work and giving a prospect of handling the tracking problem via neural fields.
光影幻象：神经辐射场中的时空流转
Neural Radiance Fields for Unconstrained Photo Collections
Introduction about Neural Radiance Fields (NeRF) for unconstrained photo collections, including NeRF, NeRF in the Wild, and Ha-NeRF.
Funding & Grants
I gratefully acknowledge support from the following programs and organizations.
- 2026: Compute Grant, MiraclePlus & Beijing Haidian Compute Lab.
奇绩创坛 & 北京海淀算力实验室算力资助 - 2025: Dean's PhD Research Grant, School of Engineering, Westlake University.
西湖大学工学院院长专项博士生项目
Teaching
Machine Learning
Westlake University
Machine Learning, Teaching Assistant, Westlake University
Academic Services
-
Area Chair
- NeurIPS
-
Outstanding Reviewer
- Computer Vision: CVPR, ICCV, ECCV, 3DV, TPAMI
- Machine Learning: NeurIPS, ICLR, ICML
- Robotics: IROS
- Graphics: SIGGRAPH, TOG, TVCG