Zakaria Laskar presents Continual Learning for Image-Based Camera Localization

On 2021-10-07 11:00:00 at
Online seminar:

For several emerging technologies such as augmented reality, autonomous driving
and robotics, visual localization is a critical component. Structured methods
directly regressing 3D scene coordinates from the input image using deep neural
networks has shown great potential on indoor datasets like 7Scenes and
12Scenes. The high accuracy on these datasets often raises the question - are
these datasets saturated and not required in future benchmarks?

In this talk, we will approach the problem of visual localization in a
continual learning setup – whereby the model is trained on scenes in an
incremental manner. This setup is more practical as new scenes often need to be
added or existing scenes undergo change requiring to retrain localization
models. Ideally, localization models should be able to adapt to new scenes
retaining performance on previously seen scenes. Our results show that existing
top-performing localization models fail to adapt to this scenario on these
so-called saturated datasets. I will present some of the standard best
practices in continual learning literature to address the continual
problem. Finally, we will see how these best practices fail to account for the
geometric information in the scene that is very important in the visual
localization problem. A very trivial approach will be presented that alleviates
this limitation and provides further improvement in results on 7Scenes,
and also 19Scenes by combining the former scenes.

Za obsah zodpovídá: Petr Pošík