FINER++: Building a Family of Variable-periodic Functions for Activating Implicit Neural Representation

Area of Science:

• Signal Processing
• Neural Networks
• Computer Vision

Background:

• Implicit Neural Representations (INRs) map coordinates to attributes, revolutionizing signal processing.
• Current INRs exhibit spectral bias and capacity-convergence gaps, limiting complex signal representation.

Purpose of the Study:

• To address limitations in current INR techniques.
• To propose a novel framework, FINER++, for enhanced signal representation.

Main Methods:

• Extending activation functions to variable-periodic ones within the FINER++ framework.
• Initializing neural network bias to activate sub-functions with diverse frequencies.
• Employing variable-periodic activation functions to flexibly tune the supported frequency set.

Main Results:

• FINER++ demonstrates improved performance in representing complex signals with multiple frequencies.
• Successful generalization across various activation function backbones (Sine, Gauss., Wavelet).
• Effective application in diverse tasks including 2D image fitting, 3D signed distance fields, and 5D neural radiance fields.

Conclusions:

• FINER++ framework effectively overcomes spectral bias and capacity-convergence issues in INRs.
• Variable-periodic activation functions offer a flexible approach to enhance signal representation capabilities.
• The proposed method significantly improves upon existing INR techniques for various applications.