3D Vision

Depth-range-free MVS network using pose embedding achieves robust and accurate 3D reconstruction.

Researchers developed a lightweight protocol to probe large vision models’ 3D physical understanding by training classifiers on model features for various scene properties (geometry, material, lightin…

CAST: a novel consistency-aware spot-guided Transformer achieves state-of-the-art accuracy and efficiency in point cloud registration.

4Diffusion generates high-quality, temporally consistent 4D content from monocular videos using a unified multi-view diffusion model and novel loss functions.

Uncertainty-aware 4D Gaussian Splatting enhances dynamic scene reconstruction from monocular videos by selectively applying regularization to uncertain regions, improving both novel view synthesis and…

3DGS-Enhancer boosts unbounded 3D Gaussian splatting, generating high-fidelity novel views even with sparse input data using view-consistent 2D diffusion priors.

3DET-Mamba: A novel end-to-end 3D object detector leveraging the Mamba state space model for efficient and accurate object detection in complex indoor scenes, outperforming previous 3DETR models.

Researchers rethink 3D Gaussian Splatting as MCMC sampling, improving rendering quality and Gaussian control via a novel relocation strategy.

Learned fragment pruning accelerates 3D Gaussian splatting rendering by selectively removing fragments, achieving up to 1.71x speedup on edge GPUs and 0.16 PSNR improvement.

3DFMNet: A novel two-stage network for multi-instance point cloud registration, achieving state-of-the-art accuracy by focusing on object centers first and then performing pairwise registration.

3D pose estimation is revolutionized by a novel SO(3)-equivariant network directly predicting Wigner-D harmonics, achieving state-of-the-art accuracy and efficiency.

SE(3)-equivariant ray embeddings in Perceiver IO achieve state-of-the-art implicit multi-view depth estimation, surpassing methods that rely on data augmentation for approximate equivariance.

Di²Pose, a novel discrete diffusion model, tackles occluded 3D human pose estimation by employing a two-stage process: pose quantization and discrete diffusion, achieving state-of-the-art results.