A Novel Entropy-Based Regularization for NeRF to View Synthesis in Few-Shot Scenarios

The Neural Radiance Field (NeRF) has marked a significant advance in 3D computer vision and graphics, offering the ability to generate high-quality views from multiple images. However, its performance is significantly limited in scenarios with sparse image data. To address this issue, our approach introduces a novel entropy-based regularization from the view of information theory, suppressing the noise in the radiance field, which is prevalent with sparse input scenarios and causes visual artifacts. By implementing our entropy-based regularization, we effectively enhance the quality of the output demonstrated by diverse experiments. Our method provides a simple yet practical, computationally efficient solution for few-shot NeRF, paving the way for NeRF in real-world applications where image data is typically sparse.