Acoustic Camera Pose Refinement Using Differentiable Rendering

Abstract

Acoustic cameras, also known as 2D forward looking sonars, show high reliability in underwater environments as they can produce high resolution images even if the illumination is limited. However, due to the unique imaging principle, it is hard to estimate ground-truth-level extrinsic parameters even in a known 3D scene. Usually, there are methods such as direct measurements by rulers to acquire a rough pose with centimeter-level error. It is necessary to refine the pose to millimeter-level error. In this work, we develop a novel differentiable acoustic camera simulator, which can be applied for estimating accurate 6 degrees of freedom pose of the acoustic cameras. We calculate the derivatives of synthetic acoustic images with respect to camera pose, and further integrated them into a gradient-based optimization pipeline to refine the pose. To mitigate the domain gap between real and synthetic images, an unpaired image translation method is used to transfer the real image to synthetic domain. Experiments prove the feasibility of the proposed method. It outperforms methods of previous research for higher efficiency and accuracy.