#3D Reconstruction

Discover advancements in converting single images to 3D models using zero-shot learning techniques. This approach mitigates the Janus problem by employing camera modeling and using synthetic datasets such as Objaverse. Recent updates, including Zero123-XL and integrations with Threestudio, improve view synthesis and 3D reconstruction. Explore live demos, optimized codebases for RTX GPUs, and training scripts. This project is a collaboration between Columbia University and Toyota Research Institute, utilizing technology from Stable Diffusion and NeRDi.

This platform presents an organized collection of 3D Gaussian Splatting research, including papers and reviews from leading conferences like CVPR2024 and ECCV2024. It highlights research across various institutions, offering insights into applications such as robotics and 3D reconstruction. The repository includes resources like code and abstracts, supporting further exploration. This centralized database helps users stay updated on advancing technologies in 3D Gaussian Splatting.

The ScanNet dataset offers over 2.5 million RGB-D views from 1500 scans, complete with 3D camera poses, surface reconstructions, and semantic segmentations. It is applicable in scene understanding tasks such as 3D object classification and semantic voxel labeling. The dataset is organized by RGB-D sequence, integrating scan data, camera poses, and semantic annotations. Tools like the ScanNet C++ Toolkit and BundleFusion enhance data handling and 3D modeling. Check the ScanNet Terms of Use for accessing the data.

Era3D offers an efficient approach to high-resolution multiview image generation, utilizing a row-wise attention mechanism. It aligns input and output images accurately, especially for specific applications, and optimizes for minimal perspective distortions. With easy installation and inference, Era3D can be downloaded from Huggingface. This open-source project under the AGPL-3.0 license invites community innovation and contribution.

OpenLRM provides open-source 3D reconstruction models that convert single images into detailed assets. The project includes regular updates, seamless integration, and compatibility with training data from Objaverse and MVImgNet. Available on Hugging Face, users can access model weights, training scripts, and a demo interface to assist in research and development within the 3D modeling field. The restructured codebase and comprehensive documentation ensure ease of use.

Discover the 'Splatter Image' project, known for its rapid single-view 3D reconstruction. It supports any object reconstruction using a model trained on Objaverse. The project includes models for six different datasets such as ShapeNet and CO3D, delivering top-tier performance. A web application is available for demonstration, while multi-GPU training and optimized algorithms remove camera pre-processing for CO3D, utilizing Gaussian Splatting for improved visualization.

Discover a cutting-edge radiance field reconstruction method that significantly enhances speed and quality. It supports half-precision training, reducing convergence time by over 100% and improving results by approximately 1.5dB. LaRa efficiently handles multi-view, text, and single-view inputs using datasets like GObjaverse. The updated scripts facilitate seamless dataset downloads and preprocessing, while intuitive training and evaluation guides ensure ease of use. This solution is perfect for applications requiring comprehensive baseline processing.