Martin Bichsel

Automatic interpolation and recognition of face images by morphing

This article presents a new method for generating an optimum image morph field, i.e. an optimum mapping of one image to another image by distorting the brightness and geometry of the former image. The mapping is calculated by maximizing the probability of the morph field in a Bayesian framework. In contrast to other techniques this new method needs no training and is derived based on group invariances and expected singularities, only. An infeasible exhaustive search is replaced by an new iterative approximation approach where, in a neighborhood around the current morph field solution, the probability distribution is approximated by a Gaussian probability distribution for which the most likely solution is evaluated by linear algebra techniques. New views are generated by applying linearly interpolated morph fields to the original reference image. Experiments demonstrate that this approach is well suited to interpolate between different views of a single face image or between images of different persons. Finally, a new face recognition algorithm makes use of the fact that morphs among images of a single person are confined to a five-dimensional subspace within the space of all possible morphs.

Illumination invariant object recognition

Varying illumination is severe problem for existing face recognition algorithms. Altering the light direction from left to right, for example, causes a change of contrast in large face regions and causes most face recognition algorithms to fail. Theoretical results, based on the law of incoherent light superposition, provide the solid ground on which a new illumination invariant recognition algorithm is derived. A face recognition experiment demonstrates that this algorithm indeed shows improved recognition performance even if the conditions, for which the theoretical results were derived, do not hold exactly.

Analyzing a Scene's Picture Set under Varying Lighting

Based on physical laws of optics, this article analytically derives a complete description of the set of all pictures that can be taken from a given scene under varying lighting, where the camera, the scene, and the light sources are static but where each light source can vary arbitrarily in radiance. It will be shown that this picture set forms a single convex region in picture space, where each picture is represented as a high-dimensional vector. An optimum radiance invariant projection is derived which is invariant under a simultaneous change of the radiances of all light sources and which decreases the dimension of the convex region by one. The theoretical predictions are confirmed experimentally for a scene illuminated with two light sources varying in intensity. A simple lighting invariant recognition algorithm is introduced and tested in a face recognition experiment. A comparison between the new algorithm and a standard recognition algorithm is presented, which shows that lighting invariant recognition leads to considerably better performance, even if the radiance of the light sourcesandthe lighting directions change.

Hierarchical Probability Estimation

Estimating probabilities based on measured numbers of occurrences of events provides a central link from probability theory to real world applications. In an important class of applications the probabilistic events correspond to the digitized outcome of an analog sensor. This paper shows theoretically and experimentally that such events are governed by a natural similarity relation which imposes subtle but highly effective a priori constraints on the meta-probability distribution, i.e. the probability distribution of probability values. The application of these constraints significantly improves the probability estimates based on the measurements. The results are applied to estimating order-1 Markov model parameters in image processing applications.

Fast Adaptive Arithmetic Coding

The number of operations in the coding part of adaptive arithmetic coding is independent of the number of symbols. The number of operations in a traditional implementation of the adaptive part, however, increases linearly with the number of symbols. therefore, the adaptive updating of the model consumes the vase majority of computational operations if the number of symbols is large, as is typical in image coding. This paper presents a fast alternative of implementing the adaptive part in a hierarchical fashion so that the number of operations depends only logarithmically on the number of symbols.