SparseSurf: Sparse-View 3D Gaussian Splatting for Surface Reconstruction

A new framework for sparse-input novel view synthesis has been introduced, focusing on test-time natural video completion. This approach utilizes pretrained video diffusion models to generate plausible in-between views from limited scene glimpses, enhancing spatial coherence and scene reconstruction through an iterative feedback loop.

A new framework named MVS-TTA has been introduced to enhance multi-view stereo (MVS) methods through test-time adaptation (TTA) via meta-auxiliary learning. This approach combines the strengths of learning-based and optimization-based methods, utilizing a self-supervised, cross-view consistency loss to improve adaptability during inference. The model-agnostic nature of MVS-TTA allows it to be applied across various MVS techniques.

SegSplat has been introduced as a novel framework that combines rapid, feed-forward 3D reconstruction with open-vocabulary semantic understanding. It constructs a compact semantic memory bank from multi-view 2D features and predicts discrete semantic indices alongside geometric attributes for each 3D Gaussian in a single pass, enhancing the efficiency of scene semantic integration.

A new method called ReCoGS has been introduced for real-time recoloring of scenes using Gaussian Splatting, which is recognized for its efficiency in novel view synthesis and high-quality reconstructions. This user-friendly pipeline allows precise selection and recoloring of regions within pre-trained scenes, demonstrating real-time performance through an interactive tool. Code for the method is available online.

A new framework called D-FCGS has been introduced to enhance the compression of dynamic 3D representations for Free-Viewpoint Videos (FVV). This innovative approach addresses the limitations of existing Gaussian Splatting methods by implementing a standardized Group-of-Frames structure and a dual prior-aware entropy model, which improves rate estimation and view-consistent fidelity.

The introduction of the 'Upsample Anything' framework offers a lightweight test-time optimization solution that enhances low-resolution features to high-resolution outputs without requiring prior training. This method addresses the limitations of existing feature upsampling techniques, which often necessitate dataset-specific retraining or complex optimization processes.