Fall 2005 Talk Descriptions

Segmentation, Shapes and Events: 3 Hierarchical approaches using one representation

Adrian Ion

Abstract: The irregular graph pyramid is known to be a very powerful representation on which computer vision algorithms focusing on different abstraction levels can rely. The dual graph, combinatorial map, and generalized map structures, which represent embeddings (in 2, 3, ..., nD) and encode the topology of entities in space, together with the contraction and removal operations make up a tool very well suited for construction of hierarchies using decisions made in a local neighborhood. The talk will address 3 research directions at PRIP, all using the above mentioned representation: the MST Pyramid segmentation, which uses simple similarity measures and internal/external contrast concepts to contract neighboring vertices of a region adjacency graph and thus produce a segmentation hierarchy, a spatio-temporal concept for event representation in cognitive vision, and the integral trees - a step in a search for a better shape representation.

Human Shape Perception: The Role of Priors

Zygmunt Pizlo

Abstract: Perception of a 3D shape based on the information in a 2D image requires the use of priors. The role of priors in perception was emphasized for the first time by Gestalt Psychologists about 100 years ago, who postulated that the percept involves simplicity constraints. A review of the most important experiments on the role of simplicity, as well as experience, in 3D shape perception from a single 2D image will be followed by a discussion of shape perception from two or more images. It will be demonstrated that binocular shape reconstruction is an ill-conditioned problem, which leads to unstable solutions, unless priors are used. A computational model of human binocular shape reconstruction will be presented and several formulations of the simplicity constraint will be discussed.

Fast population coding

Quentin Huys

Abstract: Uncertainty arises in neural computations from noisy processing elements and the formally ill-posed nature of many tasks. Taking appropriate decisions requires that uncertainty be represented and manipulated in a self-consistent manner, likely in standard cortical structures such as population codes. There is a rich literature on the capability of populations of neurons to support computations in the face of the two types of uncertainty. However, one major facet of uncertainty has received rather little attention, namely time, as in a dynamic, rapidly changing world, data is only temporarily relevant. Here, we analyse the computational consequences of encoding stimulus trajectories in the activity of populations of neurons. For a simple, instantaneous, analytically tractable encoder, we show how the correlations induced by natural, smooth stimuli lead to a decoding problem that can only be resolved by access to information that is non-local both in time and across neurons. Such encodings are computationally ruinous; we show that there is an alternative, computationally and representationally powerful, code in which each spike contributes independent information, {\em ie} is independently decodeable. [Work done with Richard Zemel, Rama Natarajan and Peter Dayan]

A Theory of Inverse Light Transport

Kyros Kutulakos

Abstract: In this talk I will consider the problem of computing and removing interreflections in photographs of real scenes. Towards this end, we introduce the problem of Inverse Light Transport: given a photograph of an unknown scene, decompose it into a sum of "n-bounce" images, where each image records the contribution of light that bounces exactly n times before reaching the camera. We prove the existence of a set of interreflection cancelation operators that enable computing each n-bounce image by multiplying the photograph by a matrix. This matrix is derived from a set of “impulse images” obtained by probing the scene with a narrow beam of light. The operators work under unknown and arbitrary illumination, and exist for scenes that have arbitrary spatially-varying BRDFs. We derive a closed-form expression for these operators in the Lambertian case and present experiments with textured and untextured Lambertian scenes that confirm our theory’s predictions. [This talk is in preparation for my talk at ICCV'05; it is joint work with Yasu Matsushita at MSR-Asia and Steve Seitz at U. Washington]

Brief Overviews of Research Interests

Vision Group Faculty

Abstract: This is an informal meeting to introduce vision faculty to new students, and to discuss plans for the fall.