Title: Attention-based gating units separate channels in neural radiance fields

Authors: Chufei Yu; Gongbing Su; Meng Yuan; Wenhao Zeng

Addresses: Department of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, China ' Department of Mechanical Engineering and Automation and Hubei Key Laboratory of Digital Textile Equipment, Wuhan Textile University, Wuhan, Hubei, China ' Department of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, China ' Department of Mechanical Engineering and Automation, Wuhan Textile University, Wuhan, Hubei, China

Abstract: We introduce a unique inductive bias to improve the reconstruction quality of Neural Radiance Fields (NeRF), NeRF employs the Fourier transform to map 3D coordinates to a high-dimensional space, enhancing the representation of high-frequency information in scenes. However, this transformation often introduces significant noise, affecting NeRF's robustness. Our approach allocates attention effectively by segregating channels within NeRF using attention-based gating units. We conducted experiments on an open-source data set to demonstrate the effectiveness of our method, which leads to significant improvements in the quality of synthesised new-view images compared to state-of-the-art methods. Notably, we achieve an average PSNR increase of 0.17 compared to the original NeRF. Furthermore, our method is implemented through a carefully designed special Multi-Layer Perceptron (MLP) architecture, ensuring compatibility with most existing NeRF-based methods.

Keywords: neural radiance fields; attention mechanism; implicit neural representation; multi-layer perceptron.

DOI: 10.1504/IJWMC.2024.142084

International Journal of Wireless and Mobile Computing, 2024 Vol.27 No.4, pp.335 - 345

Received: 03 Mar 2023
Accepted: 09 Oct 2023
Published online: 07 Oct 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article