Phase-MLP for Generalizable Implicit Neural Representations

Abstract

Implicit neural representation(INR) has lifted a climax among deep learning researchers for its powerful continuous representation. With Sine activation or Fourier position embedding, INRs overcome the problem that could not reconstruct high-frequency signals in different domains, such as voice, image or 3D shape. However, many of INR researches can merely represent one object or one scene with high-frequency information, multiple instances will bring a sharp decline of performance. In this work, we propose a newly phase-MLP which can not only recover diverse instances but also keep high quality content. Our network takes phase information which corresponding to target signal as input and combine it with encoded position signals to reconstruct the original data. Moreover, we propose a multi-level phase-MLP base on the infrastructure to retain fidelity for bigger instance while limiting the phase information in the low amount. Experimental results on public images and videos demonstrate our proposed approach outperforms the state-of-the-art methods.