DirectVoxGO++: Fast Neural Radiance Fields for Object Reconstruction

Abstract

In recent years, new reconstruction techniques based on Neural Radiance Fields (NeRFs) have created new forms to model objects instead of the traditional mesh and point cloud-based representations, allowing for more photorealistic rendering. However, these techniques were too slow to be used in practical settings, taking in the range of hours in high-end GPUs. Due to these limitations, new techniques have been created for fast reconstruction of scenes, such as DirectVoxGO. Alongside this limitation, one issue with NeRFs is that they were initially unable to separate the foreground from the background and had problems with 360 scenes until the emergence of new techniques such as NeRF++. Our method extends DirectVoxGO to allow the handling of unbounded scenes inspired by some ideas from NeRF++, adapting it to incorporate elements from a neural hashing approach employed by other works. Our technique improved photorealism compared with DirectVoxGO and Plenoxels on a subset of the LF dataset on average in at least 2%, 8%, and 8% for PSNR, SSIM, and LPIPS metrics, respectively, while also being an order of magnitude faster than NeRF++. Code will be available in https://github.com/danperazzo/dvgoplusplus.