Vojtěch Franc

An iterative algorithm learning the maximal margin classifier

Abstract A simple learning algorithm for maximal margin classifiers (also support vector machines with quadratic cost function) is proposed. We build our iterative algorithm on top of the Schlesinger–Kozinec algorithm (S–K-algorithm) from 1981 which finds a maximal margin hyperplane with a given precision for separable data. We suggest a generalization of the S–K-algorithm (i) to the non-linear case using kernel functions and (ii) for non-separable data. The requirement in memory storage is linear to the data. This property allows the proposed algorithm to be used for large training problems. The resulting algorithm is simple to implement and as the experiments showed competitive to the state-of-the-art algorithms. The implementation of the algorithm in Matlab is available. We tested the algorithm on the problem aiming at recognition poor quality numerals.

Greedy Algorithm for a Training Set Reduction in the Kernel Methods

We propose a technique for a training set approximation and its usage in kernel methods. The approach aims to represent data in a low dimensional space with possibly minimal representation error which is similar to the Principal Component Analysis (PCA). In contrast to the PCA, the basis vectors of the low dimensional space used for data representation are properly selected vectors from the training set and not as their linear combinations. The basis vectors can be selected by a simple algorithm which has low computational requirements and allows on-line processing of huge data sets. The proposed method was used to approximate training sets of the Support Vector Machines and Kernel Fisher Linear Discriminant which are known method for learning classifiers. The experiments show that the proposed approximation can significantly reduce the complexity of the found classifiers (the number of the support vectors) while retaining their accuracy.

Sequential coordinate-wise algorithm for the non-negative least squares problem

This paper contributes to the solution of the non-negative least squares problem (NNLS). The NNLS problem constitutes a substantial part of many computer vision methods and methods in other fields, too. We propose a novel sequential coordinate-wise algorithm which is easy to implement and it is able to cope with large scale problems. We also derive stopping conditions which allow to control the distance of the solution found to the optimal one in terms of the optimized objective function. The proposed algorithm showed promising performance in comparison to the projected Landweber method.

License plate character segmentation using hidden markov chains

We propose a method for segmentation of a line of characters in a noisy low resolution image of a car license plate. The Hidden Markov Chains are used to model a stochastic relation between an input image and a corresponding character segmentation. The segmentation problem is expressed as the maximum a posteriori estimation from a set of admissible segmentations. The proposed method exploits a specific prior knowledge available for the application at hand. Namely, the number of characters is known and its is also known that the characters can be segmented to sectors with equal but unknown width. The efficient algorithm for estimation based on dynamic programming is derived. The proposed method was successfully tested on data from a real life license plate recognition system.

RAVEL: an annotated corpus for training robots with audiovisual abilities

We introduce Ravel (Robots with Audiovisual Abilities), a publicly available data set which covers examples of Human Robot Interaction (HRI) scenarios. These scenarios are recorded using the audio-visual robot head POPEYE, equipped with two cameras and four microphones, two of which being plugged into the ears of a dummy head. All the recordings were performed in a standard room with no special equipment, thus providing a challenging indoor scenario. This data set provides a basis to test and benchmark methods and algorithms for audio-visual scene analysis with the ultimate goal of enabling robots to interact with people in the most natural way. The data acquisition setup, sensor calibration, data annotation and data content are fully detailed. Moreover, three examples of using the recorded data are provided, illustrating its appropriateness for carrying out a large variety of HRI experiments. The Ravel data are publicly available at: http://ravel.humavips.eu/.

Facial Landmarks Detector Learned by the Structured Output SVM

We propose a principled approach to supervised learning of facial landmarks detector based on the Deformable Part Models (DPM). We treat the task of landmarks detection as an instance of the structured output classification. To learn the parameters of the detector we use the Structured Output Support Vector Machines algorithm. The objective function of the learning algorithm is directly related to the performance of the detector and controlled by the user-defined loss function, in contrast to the previous works. Our proposed detector is real-time on a standard computer, simple to implement and easily modifiable for detection of various set of landmarks. We evaluate the performance of our detector on a challenging “Labeled Faces in the Wild” (LFW) database. The empirical results show that our detector consistently outperforms two public domain implementations based on the Active Appearance Models and the DPM. We are releasing open-source code implementing our proposed detector along with the manual annotation of seven facial landmarks for nearly all images in the LFW database.

Simple solvers for large quadratic programming tasks

This paper describes solvers for specific quadratic programming (QP) tasks. The QP tasks in question appear in numerous problems, e.g., classifier learning and probability density estimation. The QP task becomes challenging when large number of variables is to be optimized. This the case common in practice. We propose QP solvers which are simple to implement and still able to cope with problems having hundred thousands variables.

Greedy Kernel Principal Component Analysis

This contribution discusses one aspect of statistical learning and generalization. The theory of learning is very relevant to cognitive systems including cognitive vision.

MORD: multi-class classifier for ordinal regression

We show that classification rules used in ordinal regression are equivalent to a certain class of linear multi-class classifiers. This observation not only allows to design new learning algorithms for ordinal regression using existing methods for multi-class classification but it also allows to derive new models for ordinal regression. For example, one can convert learning of ordinal classifier with (almost) arbitrary loss function to a convex unconstrained risk minimization problem for which many efficient solvers exist. The established equivalence also allows to increase discriminative power of the ordinal classifier without need to use kernels by introducing a piece-wise ordinal classifier. We demonstrate advantages of the proposed models on standard benchmarks as well as in solving a real-life problem. In particular, we show that the proposed piece-wise ordinal classifier applied to visual age estimation outperforms other standard prediction models.

Learning CNNs from weakly annotated facial images

Abstract Learning of convolutional neural networks (CNNs) to perform a face recognition task requires a large set of facial images each annotated with a label to be predicted. In this paper we propose a method for learning CNNs from weakly annotated images. The weak annotation in our setting means that a pair of an attribute label and a person identity label is assigned to a set of faces automatically detected in the image. The challenge is to link the annotation with the correct face. The weakly annotated images of this type can be collected by an automated process not requiring a human labor. We formulate learning from weakly annotated images as a maximum likelihood (ML) estimation of a parametric distribution describing the weakly annotated images. The ML problem is solved by an instance of the EM algorithm which in its inner loop learns a CNN to predict attribute label from facial images. Experiments on age and gender estimation problem show that the proposed algorithm significantly outperforms the existing heuristic approach for dealing with this type of data. A practical outcome of our paper is a new annotation of the IMDB database [26] containing 300 k faces each one annotated by biological age, gender and identity labels.