Building digital models of ships from video sequences

A topic for a specialization project with subsequent Masters thesis

Proposer: Rudolf Mester, IDI

potential Co-supervisors: Annette Stahl (ITK), Kjetil Vasstein (ITK), Frank Lindseth (IDI), Andreas L Teigen (IDI)

Telegram style description:

• Input: videoclips of ships, possibly some mild manual annotation.

• Output: a complete (360 degree surround) model of the object, where invisible parts are completed by AI methods.

Motivation:

Producing ship models from real video footage of ships is obviously the most straightforward approach to obtain ship models without manually modeling them.

How

Using classical computer vision (SLAM methods) for model building is feasible (as available products and available research reports show), but to do it in a robust manner, without, or with only minimal human intervention, is hard. The plan is to create partial models from visual SLAM, using video sequences found on the internet, or acquired locally (using mobile phones or better: a stereo camera) and completing that model using AI methods.

• The analysis step (generating point clouds) will use novel methods currently under investigation in the AROS project (environment perception for underwater snake robots).
• the conversion into a texture 3D model should make use of available SW packages like Open3D.

This is a big task, so it is highly recommendable to have two students work on this task cooperatively.