Enhancements in 3D Reconstruction and Rendering

The field of 3D reconstruction and rendering is experiencing significant advancements, driven by innovative approaches to Gaussian Splatting and related techniques. A key direction is the elimination of densification steps, which traditionally slow down the reconstruction process and lead to suboptimal renderings. Novel methods are being proposed to enhance the initialization of 3DGS models, allowing for faster convergence and better preservation of scene details. Additionally, there is a growing interest in applying Gaussian Mixture Models to physics-aware compression of plasma distribution functions, enabling efficient and accurate simulations. Furthermore, researchers are exploring the use of cycle consistency regularization to improve controllable 3D generation, and developing new frameworks for stylization with disentangled priors on 3D Gaussians. Noteworthy papers include EDGS, which proposes a one-step approximation of scene geometry using triangulated pixels, and Volume Encoding Gaussians, which introduces a novel 3D Gaussian-based representation for scientific volume visualization. Other notable works include EG-Gaussian, which enhances 3D scene reconstruction using epipolar geometry and graph networks, and Metamon-GS, which improves rendering performance through variance-guided densification and light encoding.

Sources

EDGS: Eliminating Densification for Efficient Convergence of 3DGS

Volume Encoding Gaussians: Transfer Function-Agnostic 3D Gaussians for Volume Rendering

EG-Gaussian: Epipolar Geometry and Graph Network Enhanced 3D Gaussian Splatting

HoLa: B-Rep Generation using a Holistic Latent Representation

Metamon-GS: Enhancing Representability with Variance-Guided Densification and Light Encoding

VGNC: Reducing the Overfitting of Sparse-view 3DGS via Validation-guided Gaussian Number Control

IXGS-Intraoperative 3D Reconstruction from Sparse, Arbitrarily Posed Real X-rays

Physics-Aware Compression of Plasma Distribution Functions with GPU-Accelerated Gaussian Mixture Models

Cyc3D: Fine-grained Controllable 3D Generation via Cycle Consistency Regularization

StyleMe3D: Stylization with Disentangled Priors by Multiple Encoders on 3D Gaussians

Data assimilation with model errors

HUG: Hierarchical Urban Gaussian Splatting with Block-Based Reconstruction

Dense Air Pollution Estimation from Sparse in-situ Measurements and Satellite Data

Conditional Diffusion-Based Retrieval of Atmospheric CO2 from Earth Observing Spectroscopy

When Gaussian Meets Surfel: Ultra-fast High-fidelity Radiance Field Rendering

Built with on top of