Adam Breitholtz, MPENM
Understanding PoseNet - An investigation into the performance of 6DOF pose estimation neural networks
Examiner: Torsten Sattler, Dept of Electrical Engineering
Understanding the tools we use is paramount to the betterment of society and research. The popular approach of using convolutional neural networks in image analysis has resulted in undeniable successes.
However, we still have a limited understanding about how these constructs work, thus it is of great importance to decrease this gap in our knowledge. This thesis concerns itself with investigating the particular networks which are tailored towards solving the visual localization problem. The visual localization problem being the problem of finding the 6DOF camera pose from which an image was taken.
Recent work by Sattler et al. has suggested that the neural network approaches do not perform as well as one would hope and share performance characteristics with image retrieval methods. Following these results we compare two different networks with differing architecture: PoseNet and MapNet. In particular, the effects of masking parts of input images was investigated as well as the activation consistency when projecting between frames in the input sequences.
Student project presentation
Blå rummet (room 3340), Hörsalsvägen 11, 3rd floor
15 January, 2020, 13:00
15 January, 2020, 14:00