ATM-NeRF: Accelerating Training for NeRF Rendering on Mobile Devices via Geometric Regularization

Yang Chen1, Lin Zhang1, Shengjie Zhao1, Yicong Zhou2
1 School of Software Engineering, Tongji University, Shanghai, China 2 Department of Computer and Information Science, University of Macau, China
Code

Overview

Interpolate start reference image.

Overview of ATM-NeRF, which aims at accelerating training for mobile NeRF rendering via geometric regularization terms.

Demo Videos

1. The following are demo videos demonstrating the performance of our ATM-NeRF
in training acceleration on three tested datasets, respectively.




2. The following are demo videos demonstrating the rendering performance of our ATM-NeRF compared with MobileNeRF and NeRF2Mesh on iPhone XR.

"Lego" scene from the Synthetic 360 Dataset.

                            MobileNeRF                                                          NeRF2Mesh                                                          ATM-NeRF (Ours)

"Stump" scene from the Mip 360 Dataset.
MobileNeRF fails when rendering this scene on the mobilephone.

                            MobileNeRF                                                          NeRF2Mesh                                                          ATM-NeRF (Ours)

"Garden" scene from the Mip 360 Dataset.

                            MobileNeRF                                                          NeRF2Mesh                                                          ATM-NeRF (Ours)

Related Links

We thank the authors of NeRF2Mesh, Instant-NGP and nerf_pl for their public code.

To directly download our code in this page, click Code of ATM-NeRF.