Séminaire du 19/02/2018, 16h30
lundi 19 février 2018, 16h30
Robust registration with L2
Conférenciers : Rozenn Dahyot (Trinity College Dublin, School of Computer Science and Statistics)
Objective functions orginating from optimal transport and information theory frameworks are now widely used in a range of applications, from shape registration, colour transfer to machine learning.
This talk is summarising a few contributions in my team in that area, with applications to colour transfer and shape registration: probability density functions (e.g. GMMs) are used in both applications to represent colour content and shape information, and the minimisation of the Euclidian distance (L2) between two pdfs is our most recent metric for performing registration. It is shown to be robust, performant and flexible allowing to take into account correspondences when available. Recent experiment results using this approach will be shown for colour transfer and shape registration, and we will conclude with a quick overview of other techniques used in my lab (e.g. NN, MRF) as part of collaborative projects with industries.
References:
- User Interaction for Image Recolouring using L2 (2017) DOI:10.1145/3150165.3150171
- Robust Registration of Gaussian Mixtures for Colour Transfer (2017) https://arxiv.org/abs/1705.06091
- Shape Registration with Directional Data (2017) https://arxiv.org/abs/1708.07791
- Automated Colour Grading using Colour Distribution Transfer (2017) DOI:10.1016/j.cviu.2006.11.011