Learnable Camera Motion Models

This project ran 2015-2018, and was funded by The Swedish Research Council (VR).

People

The principal investigator for this project was Per-Erik Forssén.
The project also employed the PhD student Hannes Ovrén.

Research Topic

In this project, we studied continuous-time camera motion models that can be adapted to specific situations through learning. Such models are needed in video stabilisation and rectification on mobile platforms. They are also useful for control of motorised gimbals that mechanically stabilise the camera aim. Our main result is the Kontiki system that can do inertial aided 3D reconstruction from rolling shutter video. It features a novel way to balance residuals from different modalities.

This project was a continuation of the VGS project, where some early results on continuous-time representations were developed.

Software

Most software developed in this project has been released as open source.

• The Kontiki toolbox for continuous-time structure-from-motion.
• The Crisp toolbox for camera-IMU self-calibration.

Datasets

• Gopro-IMU dataset
• Gopro-Gyro dataset
• Kinect v2 dataset

Publications

• Publication list.
• DiVA publication list for the project.