Data Ingestion
We develop an easy-to-use SDK for ingesting and manipulating AV data in a scalable way with a focus on neural reconstruction applications.
Overview
The Neural Reconstruction Engine (NuRec) receives raw input data (camera/LiDAR) and outputs reconstructed scenes / assets that can be used in 3D simulation. NuRec consists of three components: data ingestion tools, a reconstruction service, and a rendering service that can be easily integrated into simulation pipelines. Multi-sensor input data is represented in the canonical NCore format that powers the ingestion and reconstruction stages.
Points of Contact:
Zan Gojcic,
Kangxue Yin
Key Research Areas
Neural Reconstruction • Feed-forward Reconstruction • Generative Priors • Simulation-ready Reconstruction • Large-scale Reconstruction
Featured Research Projects
Neural Reconstruction
Our data-driven reconstruction service enables the creation of simulation-ready scenes at scale.
Simulation
Our rendering service provides a simple API for fast and efficient rendering of reconstructed scenes in simulation pipelines.
Object Asset Reconstruction
Where the components above reconstruct entire scenes, Asset Harvester reconstructs the individual objects within them. From sparse, in-the-wild observations of vehicles, pedestrians, and riders, it produces simulation-ready 3D Gaussian assets that drop directly into NuRec scenes for object-level manipulation.
Asset Harvester
Sparse-view multiview generation lifted to feed-forward 3D Gaussian assets. Handles vehicles, pedestrians, and riders from real driving footage in seconds.