Synthetic seabeds for sonar depth estimation

The project explored how to generate virtual seabeds to train a model that estimates depth from sonar images. A synthetic scene supplies reference depths that can be paired with simulated observations during training.

  1. Generate a virtual seabed with known geometry.
  2. Pair simulated sonar observations with reference depth maps.
  3. Use the pairs to train a model to predict depth from sonar images.
Conceptual workflow for the internship project.

I co-supervised Vincent Bernardon’s master’s internship at the University of Montpellier from January to July 2026 with Nicolas Lutz and Juliette Drupt. The project brought together ICAR and LIRMM’s Marine Robotics team.

ICAR internship listing · LIRMM Marine Robotics team