3D Reconstruction & Generation
Reconstructing and generating complete, physically plausible 3D structure from unposed views.
News
- NOVA3R has been accepted to ICLR 2026!
- Back on Track has been selected as one of the best paper award candidates at ICCV 2025!
- CoProU-VO received best paper award at GCPR 2025. Congrats to Jingchao and Oussema!
- I started as a student researcher at Google XR in Zurich, working on dynamic scene modeling.
- Back on Track has been accepted to ICCV 2025 (Oral).
Show more
- AnyCam has been accepted to CVPR 2025.
- LEAP-VO was accepted to CVPR 2024.
- NeRF-SCR was accepted to ICRA 2024.
- I joined Technical University of Munich as an ELLIS PhD student.
Research
I build 3D-native systems that perceive, reconstruct, and generate dynamic worlds — the spatial foundation for world models and agents to simulate, predict, and act across space and time.
Dynamic World Modeling
Learning 4D representations and scene motion to capture how dynamic worlds evolve over time.
Spatial Perception for Agents
Localization, tracking, and motion estimation that ground agents and world models in space and time.
NOVA3R: Non-pixel-aligned Visual Transformer for Amodal 3D Reconstruction
International Conference on Learning Representations (ICLR), 2026
arXiv / Paper / Project Page / Code
A feed-forward transformer that builds a global latent representation to generate complete, amodal 3D structure from a set of unposed images.
Back on Track: Bundle Adjustment for Dynamic Scene Reconstruction
International Conference on Computer Vision (ICCV), 2025 (Oral, Best Paper Candidate)
arXiv / Paper / Project Page / Code
A method for consistent dynamic scene reconstruction via motion decoupling, bundle adjustment, and global refinement.
CoProU-VO: Combining Projected Uncertainty for End-to-End Unsupervised Monocular Visual Odometry
German Conference on Pattern Recognition (GCPR), 2025 (Oral, Best Paper Award)
arXiv / Project Page / Code
An unsupervised visual odometry method that improves pose estimation in dynamic scenes.
AnyCam: Learning to Recover Camera Poses and Intrinsics from Casual Videos
Computer Vision and Pattern Recognition Conference (CVPR), 2025
arXiv / Paper / Project Page / Code
A method for learning camera poses and intrinsics from dynamic casual videos.
DynSUP: Dynamic Gaussian Splatting from An Unposed Image Pair
Preprint, 2024
Dynamic radiance field reconstruction from only two images, enabled by object-level bundle adjustment.
LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry
Computer Vision and Pattern Recognition Conference (CVPR), 2024
arXiv / Project Page / Code / Video
A robust visual odometry system leveraging temporal context with long-term point tracking to tackle occlusions and dynamic environments.
Leveraging Neural Radiance Fields for Uncertainty-Aware Visual Localization
International Conference on Robotics and Automation (ICRA), 2024
A visual localization pipeline using rendered data from NeRF, uncertainty-guided novel view selection, and evidential scene coordinate regression.
Uncertainty-Driven Dense Two-View Structure from Motion
International Conference on Intelligent Robots and Systems (IROS), 2023
IEEE Robotics and Automation Letters (RA-L), 2023
arXiv / Project Page / Video
An accurate and reliable pipeline for dense two-view SfM using weighted bundle adjustment with robust outlier filtering and learning-based confidence modeling.
Webly Supervised Image Classification with Metadata: Automatic Noisy Label Correction via Visual-Semantic Graph
ACM International Conference on Multimedia (ACM MM), 2020 (Oral)
Webly supervised learning for semantic label confusion using visual-semantic graph with metadata-aware anchor selection and GNN-based label propagation.
Webly Supervised Image Classification with Self-Contained Confidence
European Conference on Computer Vision (ECCV), 2020
Webly supervised learning for noisy label classification via sample-wise web label correction with model confidence and pseudo machine label.
Other Projects
An Efficient and Accurate Offline Python SLAM using COLMAP
Supervised by Paul-Edouard Sarlin and Marc Pollefeys
Demo (KITTI) / Demo (Zurich) / Report
A robust and highly-extensible Python SLAM built on pycolmap; achieved better pose accuracy and significant speed improvement compared to COLMAP.
Real-time Photorealistic Neural Rendering in VR
Supervised by Sergey Prokudin
A general neural rendering pipeline for photorealistic synthesis in VR devices in real-time; demo included human neural rendering and scene style transfer.
Experiences
Google Research
Mentors: Keisuke Tateno, Federico Tombari
SenseTime Research
Mentors: Litong Feng, Wayne Zhang
Academic Services
- Conference Reviewer: CVPR, ECCV, ICCV, ICRA, IROS, ICLR, ICML, NeurIPS
- Journal Reviewer: RA-L