TS-NeRF: temporal and structural regularization for few-shot neural radiance fields

For NeRF(Neural Radiance Fields), synthesizing new views from sparse inputs poses a challenge as too few inputs can lead to artifacts in the rendered views. Recent methods have tackled this issue by introducing external supervision or utilizing regularization based on priors like depth to enhance reconstruction quality. Few-shot NeRF requires additional constraint information to ensure reconstruction quality. To address this, we employed two novel regularization methods. Firstly, we introduced a loss related to view structure, reshaping multiple random points into a batch to capture the relationships between points. This structural regularization method is termed SSLIP. Additionally, studies indicate that high-frequency signals during reconstruction hinder neural networks from effectively learning low-frequency information. Based on this research, we improved the encoding of positional codes, enabling their frequency to increase with the number of training iterations, referred to as temporal regularization. This enhancement ensures NeRF effectively learns lowfrequency information during the initial training stages. Our method, building upon the current state-of-the-art ViP-NeRF model, achieved superior results on the LLFF dataset.