Image analysis and understanding
The main issue addressed in this axis is image analysis and understanding ie. extracting useful information in digital images and videos and transfer it into relevant description and prediction models. These models provide access to high level computer vision tasks as image recognition, object detection, semantic segmentation, pose estimation, emotion recognition. They rely on image processing or machine learning and particularly deep learning.
We are currently working on specific challenges raised by real data from other scientific fields or in connection with application fields:
-
Impact of lighting on color constancy – Color constancy is the ability of the human visual system to perceive consistent colors despite variations in lighting conditions. The aim here is to imitate this behavior with vision systems.
-
Weakly supervised learning or learning using a priori knowledge – Learning representation spaces or invariant descriptors requires large volumes of annotated data. We propose methods for reducing the volume of annotated data by using self-supervision or a priori knowledge about the data.
-
Analysis of human expression and behavior – Our work aims to analyze facial expressions and body poses or movements, in particular to analyze emotions in “context-rich” scenes, or to analyze human behavior from videos.
-
Frugal and collaborative design for neural model implementation – We propose adapted and optimized methods for deploying models in cloud infrastructures, as well as in autonomous and non-autonomous Edge infrastructures. They are particularly based on distillation and federated learning.
We work in various scientific or application contexts such as: material characterization, biology, healthcare, human expression or behavior analysis, green technologies…