Photorealistic 3D scenes from video capture in under an hour
Traditional photogrammetry pipelines (COLMAP + NeRF) produce high-quality results but require hours of compute, specialized hardware, and significant manual cleanup. Clients needed fast, deployable 3D representations of physical spaces and objects for integration into mobile AR apps and web viewers. Off-the-shelf 3D scanning tools either produce low-resolution mesh output or require expensive dedicated hardware.
We implemented a 3DGS (Gaussian Splatting) reconstruction pipeline using COLMAP for structure-from-motion (SfM) to extract camera poses from multi-view input, followed by a Gaussian Splatting training loop on a GPU server. The pipeline accepts raw video from a standard smartphone, extracts frames at optimal intervals, runs SfM, trains the 3DGS model, and exports a .ply splat file. Output is served via a lightweight WebGL viewer (three.js based) that requires no plugin. The full stack is containerized: input processing runs on CPU nodes, training on GPU (NVIDIA A100/H100), output hosting on standard web infrastructure. A REST API wrapper allows clients to submit a video URL and poll for the completed scene.
Full reconstruction from a 200-frame smartphone video completes in under 45 minutes. Scenes render at 60fps in browser via WebGL on standard consumer hardware. Deployed for two real-estate clients and integrated as an optional visualization layer in the ISOTRA Designer AR app pipeline. The pipeline reduced the client's 3D content creation cost by approximately 80% compared to manual modelling.
6 weeks
We take on a small number of new engagements per quarter. If this matches your problem, let's talk.
Get in touch