A Blender scene imaged as synthetic aperture sonar

Isabelle Rathbun, a computer science PhD candidate at the University of Maryland, built the Blender scene below, and ApertureLab turns it into a physics-based synthetic aperture sonar image of the same geometry. The scene is a 100 meter square patch of seabed carrying a cluster of industrial pipe sections and a small upright object over hilly terrain. The left panel is that scene seen from above; the right panel is the simulated sonar image, formed by a HISAS-class sonar at 100 kHz over the same patch.

One scene, modeled and imaged

Two square panels side by side. Left: a shaded top-down view of a Blender scene, a hilly seabed with two long parallel pipe runs, three shorter pipe segments and a small upright object. Right: the synthetic aperture sonar image of the same scene, in blue and yellow, where slopes facing the sonar are bright and the ground behind relief is in shadow; the pipes and the small object sit in the same places as in the model.

Left: the Blender scene seen from directly above. Right: the synthetic aperture sonar image of the same 100 meter patch. The two share a footprint; along track is up and range increases to the right, with the sonar to the left of the scene. The two long pipe runs, the three shorter segments and the small object between them fall in the same places in both panels. Click the image to open a zoomable view; scroll to zoom and drag to pan.

Isabelle Rathbun built this scene. Her work studies multimodal underwater perception: pairing an optical camera with synthetic aperture sonar so that a perception system can continue to operate when the optical channel is degraded or lost. Testing that idea requires the same scene observed by both sensors with the geometry held fixed, which is what this import and imaging pipeline provides. Her profile is at linkedin.com/in/isabelle-rathbun-b42225247.

What carries from the model into the sonar image is the geometry, not the optical appearance. Acoustic scattering comes from surface shape and bottom type, so the colors and textures in the Blender file have no acoustic meaning. Because the simulator knows where every surface is, the optical and acoustic views share an exact geometry, and the ground-truth labels are produced with the image rather than annotated afterward.