murphylmf/unish

UniSH is the official code released alongside a CVPR 2026 research paper called "UniSH: Unifying Scene and Human Reconstruction in a Feed-Forward Pass." It comes from a team affiliated with HKUST and other universities, and its goal is to take a single monocular video, meaning ordinary footage from one camera, and reconstruct both the surrounding scene and any people in it in one pass through a neural network, rather than running separate models for scene geometry and human body estimation. The output includes estimated scene geometry, camera parameters, and SMPL parameters, which is a standard way of representing 3D human body shape and pose. The project is aimed at researchers and practitioners working in computer vision, particularly in human motion capture, 3D reconstruction, and related fields. It is not a general purpose consumer tool. Setting it up requires a Linux environment, a Conda environment, and either CUDA 11.8 or 12.1 for GPU acceleration. The installation process compiles several dependencies from source, including PyTorch3D, MMCV, and SAM2, using an included setup script, and it also requires downloading separate SMPL body model files and placing them in a specific folder before anything will run. This is a research codebase with real setup overhead, not a quick install. Once installed, running inference happens through a single command line call that points to an input video and an output directory, producing the reconstructed scene and human results. The repository provides links to a project page, a paper on arXiv, and hosted demos and models on Hugging Face for people who want to try the method without setting up the full pipeline locally. The code builds on and credits several prior open source projects in the 3D reconstruction and human modeling space, including GVHMR, BEDLAM, SMPL itself, VGGT, Pi3, and MoGe2. It is released under the Apache 2.0 license, a permissive license that allows reuse, modification, and commercial use as long as attribution is preserved. Given its academic origin, expect the documentation to prioritize reproducing paper results over a polished end user experience.