3D Vision

MV2Cyl: A novel method reconstructs 3D extrusion cylinder CAD models directly from multi-view images, surpassing accuracy of methods using raw 3D geometry.

A novel diffusion model reconstructs multimodal shape distributions from shading, mirroring human multistable perception.

MultiPull: a novel method reconstructing detailed 3D surfaces from raw point clouds using multi-step optimization of multi-level features, significantly improving accuracy and detail.

Ref-MC2 reconstructs high-fidelity 3D objects with inter-reflections by using a novel multi-times Monte Carlo sampling strategy, achieving superior performance in accuracy and efficiency.

Hierarchical Sinkhorn Tree (HST) robustly retrieves accurate 3D point cloud correspondences using multi-scale consistency, outperforming state-of-the-art methods.

MHCDIFF: a novel pipeline using multi-hypotheses conditioned point cloud diffusion for accurate 3D human reconstruction from occluded images, outperforming state-of-the-art methods.

MotionGS enhances deformable 3D Gaussian splatting for dynamic scenes by using motion flow to guide deformation, significantly improving reconstruction accuracy and outperforming state-of-the-art meth…

Hydra: a novel cryo-EM reconstruction method resolves both conformational and compositional heterogeneity ab initio, enabling the analysis of complex, unpurified samples with state-of-the-art accuracy…

INTEGER: a novel unsupervised point cloud registration method leveraging feature-geometry coherence for reliable pseudo-label mining and density-invariant feature learning, achieving state-of-the-art …

MIDGARD: Generate high-quality, simulatable 3D articulated assets with enhanced control and interpretability using a novel diffusion-based framework.

Humans as landmarks: A novel zero-shot monocular metric depth estimation method leverages generative models and human mesh recovery to transfer metric scale information, achieving superior generalizat…

Meta 3D AssetGen: High-quality text-to-mesh generation with realistic PBR materials and lighting, exceeding prior methods in speed and accuracy.

MeshXL: Autoregressively generating high-quality 3D meshes using a novel Neural Coordinate Field (NeurCF) representation and large language model approaches.

MeshFormer: High-quality 3D mesh generation from sparse views in seconds, using transformers and 3D convolutions.

3D Gaussian Mapping (3DGM) achieves self-supervised camera-only 3D scene decomposition by leveraging multi-traverse driving data, memorizing permanent structures while filtering out transient objects.

ManiPose: Manifold-constrained multi-hypothesis model solves 3D human pose estimation’s depth ambiguity, outperforming state-of-the-art models in pose consistency.

LuSh-NeRF: A novel model reconstructs sharp, bright NeRFs from hand-held low-light photos by sequentially modeling and removing noise and blur, outperforming existing methods.

LRM-Zero: Training large reconstruction models solely on synthetic data, achieving quality comparable to real-data trained models.

LP-3DGS learns to optimally prune 3D Gaussian splatting, achieving significant efficiency gains without compromising rendering quality via a trainable binary mask and the Gumbel-Sigmoid method.

LoD-Loc: A novel aerial visual localization method uses lightweight LoD 3D maps & neural wireframe alignment for accurate and efficient 6-DoF pose estimation, surpassing state-of-the-art methods.