Deuxième conférence plénière française de Neurosciences Computationnelles, "Neurocomp08"

8-11 octobre 2008

Marseille, France - http://2008.neurocomp.fr

Computational Vision Workshop

Neural fields are an appealing mathematical description of sets of hypercolumns and their interactions. They support the representation of such visual features as edges, textures, motion and the manner they interact in the visual field through local and far connectivity relations taking into account the variety of velocities  of  the propagation of action potentials. We motivate the neural field equations with biological considerations and go through a pedestrian tour of their main mathematical properties, existence and uniqueness of various types of solutions (stationary, homogeneous, wave like), bifurcation of these solutions with respect to some biologically relevant parameters, oscillatory behaviors. We point to some possible applications of this theory to visual illusions and voltage sensitive dyes optical imaging techniques. We also briefly relate this theory to popular techniques in computer vision, relaxation labeling, tensor voting and Bayesian belief propagation.

When you're looking for your key in your pocket, you're engaged in a process where you actively explore the shape of available objects by manipulating them with your hand until you decide that one is probably
the key you were looking for. This simple example illustrates the fact that perception is not a simple passive process where information is to be processed in order to extract knowledge. Rather, we would like to consider perception as an active exploration process where one has to decide how to handle and interact with the massive amount of information that is available anytime from the many sensors we possess. In the specific framework of vision and using distributed numerical and computational models, we would like to explain how one can actively explore a visual scene.

Binocular vision involves eye-movements, image-matching, and the interpretation of stereo-disparity information. It is typical, in computational models of stereopsis, to treat these problems quite separately. For example, it is often supposed that corresponding points should be identified in the left and right images, independently of the current orientation of the eyes. A representation of the scene is then obtained by combining separate image-based and oculomotor information. We, in contrast, emphasize the inherent geometric relationship between eye movements, binocular disparity and scene-structure. Furthermore, we show that this unified geometric treatment suggests a biologically plausible way to estimate the three-dimensional scene structure. Our analysis is based on established models from computer vision. We describe some experiments, performed with a binocular robot head, that demonstrate our approach.

Stereopsis is mainly based on horizontal disparity, especially in central vision. The implication of vertical disparity instead, present essentially in the periphery, has long been debated, mainly in the psychophysical and modelling domains. Vertical disparity can inform about relative and absolute depth. Extraretinal signals, related to eye position, are also involved in disambiguating 3D aspects of the visual scene. I will show how these different signals are encoded and interact each other at an early step of visual processing, in cortical primate area V1.

Firstly, we will present the spatio-temporal filtering, which whitens the spectrum of natural images and whose variables are not separable, a property well adapted to moving signals (an extension to neuromorphic circuits will be presented). Secondly, the coding of colour will be described and its properties of economy in wiring and energy consumption. Then, we will analyze the non-linear and adaptive functions of the retina and their properties of invariance with respect to the scene illumination. Finally, we will present the retinal sampling of photoreceptors and ganglion cells: its randomness reduces the aliasing effects and variable spatial density allows the extraction of the fundamental parameters of ego-motion.

How does a freely flying insect manage to travel safely despite the severe disturbances (obstacles, wind, etc.) it encounters all the time, given that it has no mechanical contact with the ground from which to gauge ground speed ? We will present explicit control schemes that explain how insects may navigate on the basis of optic flow cues without requiring any distance and speed measurements: how they take off and land, follow the terrain, respond suitably to headwind, avoid lateral obstacles and control their forward speed automatically. The concept of the optic flow regulator, a feedback control system based on Optic flow sensors, is presented. Optic flow sensors have been previously described in honeybees, in terms of velocity tuned neurons. We show that three OF regulators suffice to account for various insect flight patterns observed over the ground and over still water, under calm and windy conditions, and in straight and tapered corridors. To check the validity of these control schemes, they were tested in simulation and implemented onboard two types of insect-like robots, a miniature helicopter and a miniature hovercraft, which behaved very much like insects when placed in similar environments. These robots were all equipped with opto-electronic OF sensors, the underlying principle of which was based on our previous microelectrode studies on the motion sensitive neurons present in houseflies eyes. The simple and parsimonious control schemes we propose involve no conventional avionic devices such as range sensors or speed sensors. They show great potential for simplifying the design of air and space vehicles, in particular autonomous micro-aerial vehicles.

• Locate it using : Google maps
• The Campus de la Timone is at the terminus of metro Ligne 1
• Inside the Campus: Room 9 or 10 (Green aisle near the conference venur "Amphi Toga")