Gaussian See, Gaussian Do: Semantic 3D Motion Transfer from Multiview Video

The paper presents SymGS, a novel framework for compressing 3D Gaussian Splatting data by leveraging local symmetries. This approach aims to reduce the memory footprint associated with complex scenes, which can reach several gigabytes. By introducing learnable mirrors to eliminate redundant primitives, SymGS enhances existing compression methods, achieving significant improvements in efficiency compared to traditional techniques.

The introduction of 3D Gaussian Splatting (3DGS) has advanced novel view synthesis, but existing methods still face visual discrepancies at unseen sampling rates. These include blurring artifacts when zooming in and staircase artifacts when zooming out, attributed to limitations in alpha blending. A new approach, Gaussian Blending, is proposed to treat alpha and transmittance as spatially varying distributions, potentially improving rendering quality.

The paper 'Gaussian Mapping for Evolving Scenes' introduces a dynamic scene-adaptation mechanism that enhances 3D Gaussian Splatting (3DGS) for applications in computer vision, augmented reality, robotics, and autonomous driving. This mechanism continuously updates 3DGS to reflect changes in evolving scenes, addressing the limitations of existing methods that primarily focus on static scenes. A novel keyframe management system is proposed to discard outdated observations while retaining essential information.

The paper presents a novel approach to 3D Gaussian super-resolution, addressing the limitations of existing methods that operate at fixed scale factors. The proposed framework integrates scale-aware rendering, generative prior-guided optimization, and progressive super-resolving, allowing for high-resolution rendering of arbitrary scale factors using a single 3D model. Experimental results indicate a significant improvement in rendering quality, achieving a PSNR gain of 6.59 dB over traditional methods.

RoboTidy is a new benchmark designed for language-guided household tidying, addressing the limitations of current benchmarks that fail to model user preferences and support mobility. It features 500 photorealistic 3D Gaussian Splatting household scenes and provides extensive manipulation and navigation trajectories to facilitate training and evaluation in Vision-Language-Action and Vision-Language-Navigation tasks.

Sparse-view synthesis presents challenges in accurately recovering geometry and appearance from limited observations. Recent advancements in 3D Gaussian Splatting (3DGS) have improved real-time rendering quality, yet existing methods often depend on Structure-from-Motion (SfM) for camera pose estimation, which is ineffective in sparse-view scenarios. The proposed Segmentation-Driven Initialization for Gaussian Splatting (SDI-GS) addresses these inefficiencies by utilizing region-based segmentation to focus on structurally significant areas, allowing for effective downsampling of dense point cl…

The paper presents a novel approach to modeling hand-object interactions in dynamic scenes without relying on object priors. It introduces interaction-aware hand-object Gaussians with optimizable parameters to improve structural representation. The method incorporates hand shape into the object deformation field to model flexible motions and employs a progressive optimization strategy to address challenges in dynamic regions and static backgrounds.

Dental3R is a new method for intraoral 3D reconstruction that addresses the limitations of traditional intraoral scanning, which is often inaccessible for remote tele-orthodontics. Conventional techniques struggle with sparse smartphone imagery due to large view baselines, inconsistent lighting, and reflective surfaces, leading to challenges in pose and geometry estimation. Dental3R utilizes a pose-free, graph-guided approach to achieve robust, high-fidelity reconstructions from sparse intraoral photographs, overcoming issues like frequency bias that can result in loss of diagnostic details.