Photoreal Reconstruction

We present a method for photoreal scene reconstruction which uses RGB data from Aria glasses to recreate real scenes using a gaussian splatting representation.

READ THE PAPER

GITHUB REPO

DOWNLOAD THE DATASET

WATCH THE FILM

HOW DOES IT WORK?

Reconstructing scenes from video captured on any egocentric device

This novel method aims to solve the limitations of creating high-quality scene reconstruction from wearable egocentric devices. We demonstrate a solution using visual-inertial bundle adjustment (VIBA) and incorporating a physical image formation model based on Gaussian Splatting.

APPLICATIONS

Reconstructing outdoor scenes at a large scale

Using this method, we create large-scale scene reconstructions from casual RGB captures from Aria Gen 1 glasses in diverse outdoor scenarios.

Reconstructing high quality indoor scenes with lifted dynamic range

The illumination in indoor scenes can create challenges for the glasses form factor. By capturing videos with fast-exposures and visible noise, we create high dynamic range noise-free novel-view rendering.

Improving reconstruction with online calibration

The human head moves fast and frequently. We use visual inertial bundle adjustment to online calibrate RGB cameras, and use that to further model a rolling-shutter aware Gaussian Splatting reconstruction.

Creating high quality digital twin objects

We serve as a competitive baseline for high quality novel-view rendering for digital twin objects observed by Aria glasses.

Aria Scenes Dataset

Aria Scenes is a benchmark dataset for future research on photorealistic reconstruction. The dataset includes 12 .vrs files created in diverse indoor and outdoor environments.

EXPLORE THE DATASET LEARN MORE ABOUT ARIA SCENES

Related Work

In addition to Photoreal Reconstruction, related works demonstrate high quality reconstruction results on egocentric data from Project Aria glasses.

Egocentric Sparse Reconstruction Neural-PBIR

Learn more about Photoreal Reconstruction

For more information about our reconstruction method, check out our paper on arXiv.

READ THE PAPER

BibTex Citation

If you use the Aria Digital Twin Catalog in your research, please cite the following:

@inproceedings{lv2025egosplats,
    title = {Photoreal Scene Reconstruction from an Egocentric Device},
    author = {Lv, Zhaoyang and Monge, Maurizio and Chen, Ka and Zhu, Yufeng and Goesele, Michael and Engel, Jakob and Dong, Zhao and Newcombe, Richard},
    booktitle = {ACM SIGGRAPH},
    year = {2025}
}

Access Aria Scenes Dataset

If you are a researcher interested in recreating our method of scene reconstruction, access the Aria Scene Reconstruction Dataset to get started.

ACCESS THE DATASETS

By submitting your email and accessing the Aria Scene Reconstruction Dataset, you agree to abide by the dataset license agreement.