3D reconstruction and video analytics for subsea exploration

Humans are inherently visual creatures. Despite the limitations of underwater vision, video remains a critical modality for observing and interacting with the deep ocean. This talk will cover two major research themes: The first theme will present progress towards robust reconstruction of 3D environments from monocular and stereo cameras. I will show results of 3D reconstruction from ad hoc ROV video data and from camera systems developed in-house, as well as early results in fusing visual data with range information from an imaging sonar.

The second theme will involve a discussion of analytic tools for extracting data products from imagery and video sources, including geometric manipulations (e.g., image stitching and panorama generation), target tracking, and the application of machine learning techniques to challenging or repetitive data extraction problems. I will show preliminary results using machine learning to identify fauna in benthic imagery, discuss ongoing efforts to improve access to large video archives for scientific analysis, and show how to use post-processed products (panoramas and virtual reality) to engage the public with ocean science.

Speaker: Aaron Marburg, Univ. of Washington