Neural Implicit Surface Reconstruction using Imaging Sonar

Abstract

We present a technique for dense 3D reconstruction of objects using an imaging sonar, also known as forward-looking sonar (FLS). Compared to previous methods that model the scene geometry as point clouds or volumetric grids, we represent the geometry as a neural implicit function. Additionally, given such a representation, we use a differentiable volumetric renderer that models the propagation of acoustic waves to synthesize imaging sonar measurements. We perform experiments on real and synthetic datasets and show that our algorithm reconstructs high-fidelity surface geometry from multi-view FLS images at much higher quality than was possible with previous techniques and without suffering from their associated memory overhead.

Learning Pseudo Front Depth for 2D Forward-Looking Sonar-based Multi-view Stereo

“…Westman et al utilized non-line-of-sight for 3D reconstruction [16], [17]. Qadri et al proposed an implicit neural representation for surface reconstruction [6]. The aforementioned methods achieve convincing results but require numerous viewpoints.…”

“…yamashita@robot.t.u-tokyo.ac.jp images to generate 3D models. They usually require a large number of viewpoints, making it necessary to hover around the target using underwater vehicles [5], [6] or rotate the camera with motors [2]. Single-view methods utilize shadow information or model ultrasound propagation for 3D reconstructions [7].…”

Circular SAR Incoherent 3D Imaging with a NeRF-Inspired Method

“…So far, some methods based on NeRF transformations have achieved remarkable results in the problems that have close similarities to the CSAR inverse imaging problems raised in Section 2. These include extending NeRF to solve sparse-view computed tomography (CT) reconstruction problems [29], 3D imaging problems with imaging sonar data [30], and non-line-of-sight (NLOS) imaging problems with 1D light transient collections [31], among others. As the image formation model of SAR proposed in Section 2 is similar to that of imaging sonar, CT, and time-of-flight (ToF)-based NLOS imaging problems, we shall solve the non-convex optimization problem of CSAR 4D inverse imaging by adapting the NeRF model.…”