Neural Radiation Fields via Accelerated and High Quality Parallel for Novel View Synthesis

Neural Radiance Fields (NeRF) have performed well in related tasks since its introduction. However, it is often computationally costly and difficult to acquire high quality rendering. In this work, we propose an accelerated and high quality parallel approach (AHQP-NeRF) that enables neural radiation field training acceleration and improved rendering quality. We utilize multi-resolution hash coding to extract image features and apply miniaturized MLP for inference computation. In the coarse sampling stage, we sample the empty region that skipped the early termination of light in the blocked region according to the concept of transmission estimation. In the fine sampling stage, we sample using the piecewise exponential function fitting the sampling points. Our sampling approach is to speed up model training by reducing the number of points per line and combining with fine sampling to improve rendering quality. We conducted some investigations and comparative analysis of multiple SOTA methods. The results demonstrate that our method exhibits exceptional accuracy in terms of PSNR metrics when compared to other approaches. Moreover, our proposed technique significantly reduces the training time required by the NeRF method by a factor of 7.2 and we have successfully demonstrated that our approach can be trained on low-consumption devices.