Virtual LA

[Video Placeholder: Live demo navigating landmarks]

How It Works

Data collection is performed using a multi-device setup involving 360-degree cameras, DJI Osmo Pockets, and DJI drones. This hybrid approach ensures complete spatial coverage: 360 cameras capture panoramic ground-level views, DJI Osmo pockets record smooth video trajectories, and drones provide comprehensive aerial data.

These inputs are processed using state-of-the-art feature matching and structure-from-motion (SfM) techniques for accurate camera calibration and alignment, which form the backbone of high-fidelity 3D scene reconstruction.

Left: Devices Used for Data Collection; Right: Accurate 3D reconstruction and photorealistic renderings.

Rendering Technology

At the core of our real-time innovation is the adoption of 3D Gaussian Splatting (3DGS) and its integration with generative models. This combination offers speed and visual fidelity far beyond traditional rendering methods.

To support free-form exploration, we are integrating diffusion-based view synthesis models that generate plausible novel views, overcoming the limitations of fixed-camera data capture and enabling users to roam dynamically within the virtual environment.

Generative models are able to improve the rendering performance of 3DGS.

Community Platform

Our platform allows users to upload their own captured data, contributing to an ever-expanding digital urban environment. Collaborate, share, and discover new perspectives on cities worldwide.

In addition to real-time interaction, the system also serves as a large-scale dataset resource for 3D vision and generative AI research, bridging practical deployment with academic inquiry.

[Image Placeholder: Platform interface]