Article ID: 22.20250141
Recently, research on neural light field (NLF), which applies implicit neural representation (INR) to light field (LF), has been actively conducted. NLF can reconstruct dense and realistic LF from relatively sparse and unstructured images, which alleviates the high acquisition difficulty of existing LFs. On the other hand, NLF has a slow rendering speed due to pixel-level MLP processing, making real-time rendering challenging. To address real-time rendering of NLF, this paper considers the application of an explicit voxel grid (VG) data structure, which is used to improve the rendering speed of INR. In particular, the performance is compared based on the dimensions of VG. Experimental results show that the dimensions of VG involve a trade-off between rendering quality, memory usage, and training speed. The analysis presented in this paper is expected to help select the appropriate dimensions of VG according to the specific application scenario.
