ACFNeRF: Accelerating and Cache-Free Neural Rendering via Point Cloud-Based Distance Fields

Neural radiance fields offer a remarkable avenue for realistic scene rendering and novel view synthesis. Nevertheless, challenges such as sluggish training times, protracted inference durations, and limitations in handling large-scale scenes persist. To address the bottleneck of slow inference in NeRF, our propose ACFNeRF leveraging point cloud to train a distance field, improving NeRF's sampling strategy, and substantially bolstering its inference speed. Our approach achieves an impressive inference rate of 150 frames per second, enabling real-time rendering within room-scale scenes. Comprehensive experimentation validates our method's superiority, demonstrating a notable 10-20x acceleration over existing NeRF acceleration techniques under cache-free conditions.