From local descriptors to holistic models: A general framework embedding video processing and object representation

Summary form only given. We present in this talk recent and ongoing works on feature space based image processing and representation models. We choose the multiscale partial derivatives (the local jet) as a basic description of the local geometry. The local jet feature space allows local categorisation and provides a similarity based metrics which can be used in many image processing and vision tasks, including: non local denoising, inpainting, optical flow, background estimation and tracking. Furthermore, global geometrical or statistical measures in the feature space provide relevant descriptors of objects, textures, scenes or activities, that revisit from a new perspective several classic themes of computer vision, amongst which: statistical salient structures, dense Hough transforms. Finally, we consider the active vision point of view, assuming that the scene is partially and progressively acquired by a moving sensor, and investigate the model associated to this sparse and dynamic representation.