Machine Learning for Computer Vision Applications

We develop specific machine learning approaches for computer vision applications. In particular, we focus on finding relevant representations for describing images or videos, instance-based scene labeling, color constancy.

A lot of computer vision applications require the development of efficient machine learning methods. We propose to contribute in this field by developing new methods able to find better representation for images and videos for specific applications such as: Instance-based scene labeling, color constancy, perceptual distance computation, ...

We are currently interested in the following issues:

-How to take into account some background knowledge? In particular, we focus on the problem of using contextual information for improving complete scene labeling. Indeed, taking account some useful additional information on the context of the image can provide relevant features to improve labeling systems. An an example, we propose to take into account some depth information for 3D reconstruction of complete scenes. We also investigate the definition of constraints between hidden layers of neural networks. We apply our approaches to color constancy or instance-base scene labeling applications.

-How to deal with temporal information for video classification? For video classification tasks, it is essential to model recurrent temporal information occuring in the videos. We work on representation methods able to describe videos with some discriminative temporal patters to improve classification tasks. For this purpose, we work on the integration of temporal information in neural networks and the use of temporal topic models.

The research made in this area are done in collaboration with the image science team.