Geo4D: Leveraging Video Generators for Geometric 4D Scene Reconstruction
ICCV, 2025
A method to repurpose video diffusion models for monocular 3D reconstruction of dynamic scenes.
MSCA Fellowship
SYN3D: Synthesizing Photorealistic 3D Scene from Zero to One or Limited Views
MSCA Postdoctoral Fellowship (UKRI Horizon Europe)
The grant was awarded on 13/02/2024 (£ 206,085.62), and started on 01/06/2024 and will end on 01/06/2026.
Objectives
The aim of SYN3D is to enable the effortless creation of novel view synthesis and 3D reconstruction from limited views (from 0 to 5) in a feed-forward manner. In general, such a few input views do not contain sufficient information for compressive 3D synthesis. However, our scientific hypothesis is that 3D synthesis from limited views is still possible by learning, from billion images, the capability of hallucinating unobserved parts of 3D objects/scenes, thus obtaining plausible if not faithful reconstructions.
To achieve the objectives of SYN3D, the project focuses on the following key objectives:
Selected Results
MVSplat, MVSplat-360Feed-forward 3D scene synthesis from limited views (even for 360 degree)
Free3DFeed-forward 3D object synthesis from a single image for open-set categories
Publications
2025
arxivDSO: Aligning 3D Generators with Simulation Feedback for Physical Soundness
ICCV, 2025
A feed-forward approach to increase the likelihood that the 3D generator outputs stable 3D objects directly.
Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images
ICCV, 2025
A modal for completed 3D reconstruction from partially visible inputs.
Puppet-Master: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamics
ICCV, 2025
An interactive video generative model that can serve as a motion prior for part-level dynamics.
Flash3D: Feed-Forward Generalisable 3D Scene Reconstruction from a Single Image
3DV, 2025
Stanislaw Szymanowicz*, Eldar Insafutdinov*, Chuanxia Zheng*, Dylan Campbell, João Henriques, Christian Rupprecht, Andrea Vedaldi
A fast, super efficient, trainable on a single GPU in one day for scene 3D reconstruction from a single image.
2024
Splatt3R: Zero-shot Gaussian Splatting from Uncalibrated Image Pairs
arXiv, 2024
A Feed-Forward model that reconstruct 3D structure and apperance with uncalibrated images.
MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views
NeurIPS, 2024
A feed-forward approach for 360-degree scene-level novel view synthesis using only sparse observations.
MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images
ECCV, 2024(oral)
Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, Jianfei Cai
A feed-forward approach for efficiently predicting 3D Gaussians from sparse multi-view images in a single forward pass.
DragAPart: Learning a Part-Level Motion Prior for Articulated Objects
ECCV, 2024
A physical interaction with objects in vision for part-level dragging.
ClusteringSDF: Self-Organized Neural Implicit Surfaces for 3D Decomposition
ECCV, 2024
A self-organized 3D segmentation/decomposition model via neural implicit surface representation.
Free3D: Consistent Novel View Synthesis without 3D Representation
CVPR, 2024
A method to synthesize consistent novel views from a single image on open-set categories without the need of explicit 3D representations.