Orateur : Nicolas Brodu
Affiliation : Inria Bordeaux, équipe GEOSTAT
Titre : Super-resolution of multispectral images and Stochastic image analysis
The first part of this talk concerns a new resolution enhancement model, with application to MODIS and Sentinel-2 data. A mixture model is fit at high-resolution bands and separates pixel content from color information. This sub-pixellic model is then applied to unmix low-resolution bands. This technique drastically improves the overall land cover classification accuracy, especially in the MODIS cases, and it is particularly effective on Sentinel-2 data.
The second part of this talk concerns a new low-level boundary detector. The method relies on constrained random walks around each pixel, describing how nearby image values typically evolve on each side of this pixel. Textures are represented as probability distributions of such random walks, and a texture difference operator is statistically defined as a distance between these distributions in a suitable reproducing kernel Hilbert space. The method is thus not limited to scalar pixel values: any data type for which a kernel is available may be considered. The method is implicitly scale-dependent and can also be used to infer characteristic scales in measured data.
Orateur : Carlos F. CRISPIM-JUNIOR
Affiliation : INRIA Sophia Antipolis
Titre : Multimodal scene understanding and its medical applications
Résumé : The task of visual understanding of complex scenes, more specifically activity recognition, has been investigated for several years. But, most existing work still refers to action recognition methods over brief, pre-clipped, single action videos. My research project goes beyond the state of the art and targets the recognition of complex human activities from multiple modalities and sensors (eg., wearable and 3D cameras). I seek to decompose complex activities in deep semantic models, where atomic semantic units are called concepts and are used to abstract semantics from low-level data. From the conceptual representation level, I investigate different forms to determine concept relations: from domain expert knowledge up to machine learning methods.
My presentation will target three subjects: recognition of activities of daily living using a knowledge-driven approach and 3D sensing (e,g., Kinect, Asus Xtion Live PRO sensors), Multimodal Activity Recognition (Kinect, GoPro, video camera), and the application of such methods to study Alzheimer's disease.