PBVS Workshop · CVPR 2026
Single-View Seafloor Recovery from Imaging Sonar via Differentiable Rendering
Sevan Brodjian · Michael Hobley · Pietro Perona
California Institute of Technology
Abstract
Forward-looking sonar (FLS) is often the only imaging modality available underwater. Each frame collapses vertical structure into a flat range-azimuth image, leaving scene elevation ambiguous. Existing 3D recovery pipelines typically require many views, multi-sensor rigs, or large quantities of labeled training data.
We present a differentiable rendering system for forward-looking imaging sonar. The renderer models the full acquisition physics: acoustic ray casting through a 3D scene, beam geometry, surface reflectance, Gaussian range binning, and log-amplitude compression, all differentiable. This makes the system usable as a component in any gradient-based optimization or learning pipeline. We demonstrate it on recovering riverbed and seafloor geometry from a single sonar frame with no training data. Scene geometry is parameterized as an explicit height field and gradient descent drives the simulated image to match the real sensor reading. The system is grounded in real sensor parameters and transfers across hardware and environments without modification.