Neural Density-Distance Fields

  • 2022-07-29 04:13:25
  • Itsuki Ueda, Yoshihiro Fukuhara, Hirokatsu Kataoka, Hiroaki Aizawa, Hidehiko Shishido, Itaru Kitahara
  • 37

Abstract

The success of neural fields for 3D vision tasks is now indisputable.Following this trend, several methods aiming for visual localization (e.g.,SLAM) have been proposed to estimate distance or density fields using neuralfields. However, it is difficult to achieve high localization performance byonly density fields-based methods such as Neural Radiance Field (NeRF) sincethey do not provide density gradient in most empty regions. On the other hand,distance field-based methods such as Neural Implicit Surface (NeuS) havelimitations in objects' surface shapes. This paper proposes NeuralDensity-Distance Field (NeDDF), a novel 3D representation that reciprocallyconstrains the distance and density fields. We extend distance fieldformulation to shapes with no explicit boundary surface, such as fur or smoke,which enable explicit conversion from distance field to density field.Consistent distance and density fields realized by explicit conversion enableboth robustness to initial values and high-quality registration. Furthermore,the consistency between fields allows fast convergence from sparse pointclouds. Experiments show that NeDDF can achieve high localization performancewhile providing comparable results to NeRF on novel view synthesis. The code isavailable at https://github.com/ueda0319/neddf.

 

Quick Read (beta)

loading the full paper ...