3dreconstruction

[arxiv'26] GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation

[ECCV 2026] Track4World: Feedforward World-centric Dense 3D Tracking of All Pixels

The official implementation of the NeurIPS 2024 paper: DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus

List of recent advances for human avatars, including generation, reconstruction, and editing, etc.

We propose a framework that accurately derives the 3D tongue shape from single images. A high detailed 3D point cloud of the tongue surface and a full head topology along with the tongue expression can be estimated from the image domain.

Official code for ICCV 2023 paper: "TransHuman: A Transformer-based Human Representation for Generalizable Neural Human Rendering".

PyTorchGeoNodes is a PyTorch module for differentiable shape programs / procedural models in forms of graphs. It can automatically translate Blender geometry node models into PyTorch code. Originally, it was designed to simplify the integration of procedural shape programs into machine learning pipelines for 3D scene understanding.

(ECCV2026 MLP Splatting) Each mlp is a light field primitive instead Gaussians. We splat mlps, making the representation more expressive, fewer primitive numbers, and align with object semantics.

ResBIM dataset — Fully Automated Synthetic BIM Dataset Generation

Master thesis work on 3d object reconstruction from 2d images with graph convolution network (pixel2mesh adaption).

[NeurIPS 2026] Click3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks

TC3DGS: Temporally Compressed 3D Gaussian Splatting for Dynamic Scenes

Demos and results of our method. Code will be available soon...

This repository showcases our project, presenting an innovative approach to 3D Indoor Mapping and Object Segmentation. With a primary focus on robot navigation in complex environments, we introduce a methodology that uses RGB images for mapping and object segmentation by integrating SimpleRecon and Point-Voxel CNN for efficient scene reconstruction

Open Museum is a project dedicated to capturing and maintaining 3D models of monuments, public art and cultural heritage globally using 3D reconstruction.

[BMVC 2026] Feed-forward 3D Gaussian splatting for uncalibrated object-centric novel view synthesis, using jointly trained VGGT geometry and query-based multi-view Gaussian decoding.

Multi-scale AANet+ using iterative residuals and warped features for real-time and scalable state-of-the-art stereo matching.

[SIGGRAPH Asia'2026] X-Splat: Unified Gaussian Splatting for Joint RGB and X-ray 3D Reconstruction

(END-VO) Endoscope Localization and Dense Surgical Scene Reconstruction for Stereo Endoscopy by Unsupervised Optical Flow and Kanade-Lucas-Tomasi Tracking

3D Reconstruction, Viewer3D, Volumetric renderer, Differential rendering, tool to experiment with 3D and shapes.