Showing papers in "Pattern Recognition Letters in 2008"
TL;DR: This paper introduces k'-means algorithm that performs correct clustering without pre-assigning the exact number of clusters by minimizing a suggested cost-function.
Abstract: This paper introduces k'-means algorithm that performs correct clustering without pre-assigning the exact number of clusters. This is achieved by minimizing a suggested cost-function. The cost-function extends the mean-square-error cost-function of k-means. The algorithm consists of two separate steps. The first is a pre-processing procedure that performs initial clustering and assigns at least one seed point to each cluster. During the second step, the seed-points are adjusted to minimize the cost-function. The algorithm automatically penalizes any possible winning chances for all rival seed-points in subsequent iterations. When the cost-function reaches a global minimum, the correct number of clusters is determined and the remaining seed points are located near the centres of actual clusters. The simulated experiments described in this paper confirm good performance of the proposed algorithm.
427 citations
TL;DR: This paper presents a systematic design approach for constructing neural classifiers that are capable of classifying human activities using a triaxial accelerometer and adopts neural networks as the classifiers for activity recognition.
Abstract: This paper presents a systematic design approach for constructing neural classifiers that are capable of classifying human activities using a triaxial accelerometer. The philosophy of our design approach is to apply a divide-and-conquer strategy that separates dynamic activities from static activities preliminarily and recognizes these two different types of activities separately. Since multilayer neural networks can generate complex discriminating surfaces for recognition problems, we adopt neural networks as the classifiers for activity recognition. An effective feature subset selection approach has been developed to determine significant feature subsets and compact classifier structures with satisfactory accuracy. Experimental results have successfully validated the effectiveness of the proposed recognition scheme.
365 citations
TL;DR: A multi-purpose image classifier that can be applied to a wide variety of image classification tasks without modifications or fine-tuning, and yet provide classification accuracy comparable to state-of-the-art task-specific image classifiers.
Abstract: We describe a multi-purpose image classifier that can be applied to a wide variety of image classification tasks without modifications or fine-tuning, and yet provide classification accuracy comparable to state-of-the-art task-specific image classifiers. The proposed image classifier first extracts a large set of 1025 image features including polynomial decompositions, high contrast features, pixel statistics, and textures. These features are computed on the raw image, transforms of the image, and transforms of transforms of the image. The feature values are then used to classify test images into a set of pre-defined image classes. This classifier was tested on several different problems including biological image classification and face recognition. Although we cannot make a claim of universality, our experimental results show that this classifier performs as well or better than classifiers developed specifically for these image classification tasks. Our classifier's high performance on a variety of classification problems is attributed to (i) a large set of features extracted from images; and (ii) an effective feature selection and weighting algorithm sensitive to specific image classification problems. The algorithms are available for free download from http://www.openmicroscopy.org.
280 citations
TL;DR: A novel algorithm is proposed for segmenting an image into multiple levels using its mean and variance, making use of the fact that a number of distributions tend towards Dirac delta function, peaking at the mean, in the limiting condition of vanishing variance.
Abstract: A novel algorithm is proposed for segmenting an image into multiple levels using its mean and variance. Starting from the extreme pixel values at both ends of the histogram plot, the algorithm is applied recursively on sub-ranges computed from the previous step, so as to find a threshold level and a new sub-range for the next step, until no significant improvement in image quality can be achieved. The method makes use of the fact that a number of distributions tend towards Dirac delta function, peaking at the mean, in the limiting condition of vanishing variance. The procedure naturally provides for variable size segmentation with bigger blocks near the extreme pixel values and finer divisions around the mean or other chosen value for better visualization. Experiments on a variety of images show that the new algorithm effectively segments the image in computationally very less time.
249 citations
TL;DR: A new face recognition algorithm based on the well-known EBGM which replaces Gabor features by HOG descriptors is presented which shows a better performance compared to other face recognition approaches using public available databases.
Abstract: This paper presents a new face recognition algorithm based on the well-known EBGM which replaces Gabor features by HOG descriptors. The recognition results show a better performance of our approach compared to other face recognition approaches using public available databases. This better performance is explained by the properties of HOG descriptors which are more robust to changes in illumination, rotation and small displacements, and to the higher accuracy of the face graphs obtained compared to classical Gabor-EBGM ones.
229 citations
TL;DR: An accumulative motion model based on the integral image by fast estimating the motion orientation of smoke is proposed, which together with chrominance detection can correctly detect the existence of smoke.
Abstract: Video smoke detection has many advantages over traditional methods, such as fast response, non-contact, and so on. But most of video smoke detection systems usually have high false alarms. In order to improve the performance of video smoke detection, we propose an accumulative motion model based on the integral image by fast estimating the motion orientation of smoke. But the estimation is not very precise due to block sum. Not very accurate estimation will affect the subsequent decision. To reduce this influence, the accumulation of the orientation over time is performed to compensate results for the inaccuracy of orientation. The model is able to mostly eliminate the disturbance of artificial lights and non-smoke moving objects by using the accumulation of motion. The model together with chrominance detection can correctly detect the existence of smoke. Experimental results show that our algorithm has good robustness for smoke detection.
222 citations
TL;DR: A self-training semi-supervised support vector machine (SVM) algorithm and its corresponding model selection method, which are designed to train a classifier with small training data are presented and the convergence of this algorithm is proved.
Abstract: In this paper, we first present a self-training semi-supervised support vector machine (SVM) algorithm and its corresponding model selection method, which are designed to train a classifier with small training data. Next, we prove the convergence of this algorithm. Two examples are presented to demonstrate the validity of our algorithm with model selection. Finally, we apply our algorithm to a data set collected from a P300-based brain computer interface (BCI) speller. This algorithm is shown to be able to significantly reduce training effort of the P300-based BCI speller.
221 citations
TL;DR: The main contributions are comprehensive and comparable classification results for the gender classification methods combined with automatic real-time face detection and, in addition, with manual face normalization, and two new variants of the known methods.
Abstract: Successful face analysis requires robust methods. It has been hard to compare the methods due to different experimental setups. We carried out a comparison study for the state-of-the-art gender classification methods to find out their actual reliability. The main contributions are comprehensive and comparable classification results for the gender classification methods combined with automatic real-time face detection and, in addition, with manual face normalization. We also experimented by combining gender classifier outputs arithmetically. This lead to increased classification accuracies. Furthermore, we contribute guidelines to carry out classification experiments, knowledge on the strengths and weaknesses of the gender classification methods, and two new variants of the known methods.
203 citations
TL;DR: A multifocus image fusion algorithm based on combination of wavelet and curvelet transform is proposed, which exhibits clear advantages over any individual transform alone.
Abstract: When an image is captured by CCD device, only the objects at focus plane would appear sharp. A practicable way to get an image with all objects in focus is to fuse images acquired with different focus levels of the scene. In this paper, we propose a multifocus image fusion algorithm based on combination of wavelet and curvelet transform. Although the fused results obtained by wavelet or curvelet transform individually are encouraging, there is still large room for further improvement because wavelets do not represent long edges well while curvelets are challenged with small features. So in the proposed method, these two methods are combined together. Each of the registered images is decomposed using curvelet transform firstly. Then the coefficients are fused using wavelet-based image fusion method. Finally, the fused image is reconstructed by performing the inverse curvelet transform. The experimental results on several images show that the combined fusion algorithm exhibits clear advantages over any individual transform alone.
200 citations
TL;DR: A scheme for clustering complex and linearly non-separable datasets, without any prior knowledge of the number of naturally occurring groups in the data is introduced, based on a modified version of classical Particle Swarm Optimization algorithm, known as the Multi-Elitist PSO (MEPSO) model, which employs a kernel-induced similarity measure instead of the conventional sum-of-squares distance.
Abstract: This article introduces a scheme for clustering complex and linearly non-separable datasets, without any prior knowledge of the number of naturally occurring groups in the data. The proposed method is based on a modified version of classical Particle Swarm Optimization (PSO) algorithm, known as the Multi-Elitist PSO (MEPSO) model. It also employs a kernel-induced similarity measure instead of the conventional sum-of-squares distance. Use of the kernel function makes it possible to cluster data that is linearly non-separable in the original input space into homogeneous groups in a transformed high-dimensional feature space. A new particle representation scheme has been adopted for selecting the optimal number of clusters from several possible choices. The performance of the proposed method has been extensively compared with a few state of the art clustering techniques over a test suit of several artificial and real life datasets. Based on the computer simulations, some empirical guidelines have been provided for selecting the suitable parameters of the PSO algorithm.
187 citations
TL;DR: This work addresses the often overlooked fact that the uncertainty in a point estimate obtained with CV and BTS is unknown and quite large for small sample classification problems encountered in biomedical applications and elsewhere and suggests that the final classification performance should be reported in the form of a Bayesian confidence interval obtained from a simple holdout test or using some other method that yields conservative measures of the uncertainty.
Abstract: The interest in statistical classification for critical applications such as diagnoses of patient samples based on supervised learning is rapidly growing. To gain acceptance in applications where the subsequent decisions have serious consequences, e.g. choice of cancer therapy, any such decision support system must come with a reliable performance estimate. Tailored for small sample problems, cross-validation (CV) and bootstrapping (BTS) have been the most commonly used methods to determine such estimates in virtually all branches of science for the last 20 years. Here, we address the often overlooked fact that the uncertainty in a point estimate obtained with CV and BTS is unknown and quite large for small sample classification problems encountered in biomedical applications and elsewhere. To avoid this fundamental problem of employing CV and BTS, until improved alternatives have been established, we suggest that the final classification performance always should be reported in the form of a Bayesian confidence interval obtained from a simple holdout test or using some other method that yields conservative measures of the uncertainty.
TL;DR: This paper attempts to solve the primal quadratic programming problems of twin SVM by converting them into smooth unconstrained minimization problems using the well-known Newton-Armijo algorithm.
Abstract: This paper enhances the recently proposed twin SVM Jayadeva et al. [Jayadeva, Khemchandani, R., Chandra, S., 2007. Twin support vector machines for pattern classification. IEEE Trans. Pattern Anal. Machine Intell. 29 (5), 905-910] using smoothing techniques to smooth twin SVM for binary classification. We attempt to solve the primal quadratic programming problems of twin SVM by converting them into smooth unconstrained minimization problems. The smooth reformulations are solved using the well-known Newton-Armijo algorithm. The effectiveness of the enhanced method is demonstrated by experimental results on available benchmark datasets.
TL;DR: A new approach based on ant colony optimization (ACO) for attribute reduction in rough set theory is introduced and it is demonstrated that this algorithm can provide competitive solutions efficiently.
Abstract: Attribute reduction in rough set theory is an important feature selection method. Since attribute reduction is an NP-hard problem, it is necessary to investigate fast and effective approximate algorithms. In this paper, we introduce a new approach based on ant colony optimization (ACO) for attribute reduction. To verify the proposed algorithm, numerical experiments are carried out on thirteen small or medium-sized datasets and three gene expression datasets. The results demonstrate that this algorithm can provide competitive solutions efficiently.
TL;DR: This paper presents a Gaussian kernel-based fuzzy c-means algorithm (GKFCM) with a spatial bias correction that becomes a generalized type of FCM, BCFCM, KFCM_S"1 and KFCS"2 algorithms and presents with more efficiency and robustness.
Abstract: Bias-corrected fuzzy c-means (BCFCM) algorithm with spatial information is especially effective in image segmentation. Since it is computationally time taking and lacks enough robustness to noise and outliers, some kernel versions of FCM with spatial constraints, such as KFCM_S"1 and KFCM_S"2, were proposed to solve those drawbacks of BCFCM. However, KFCM_S"1 and KFCM_S"2 are heavily affected by their parameters. In this paper, we present a Gaussian kernel-based fuzzy c-means algorithm (GKFCM) with a spatial bias correction. The proposed GKFCM algorithm becomes a generalized type of FCM, BCFCM, KFCM_S"1 and KFCM_S"2 algorithms and presents with more efficiency and robustness. Some numerical and image experiments are performed to assess the performance of GKFCM in comparison with FCM, BCFCM, KFCM_S"1 and KFCM_S"2. Experimental results show that the proposed GKFCM has better performance.
TL;DR: Performance of this new algorithm is compared to other popular approaches such as MADAM ID and 3-level tree classifiers, and significant improvement has been achieved from the viewpoint of both high intrusion detection rate and reasonably low false alarm rate.
Abstract: With increasing connectivity between computers, the need to keep networks secure progressively becomes more vital. Intrusion detection systems (IDS) have become an essential component of computer security to supplement existing defenses. This paper proposes a multiple-level hybrid classifier, a novel intrusion detection system, which combines the supervised tree classifiers and unsupervised Bayesian clustering to detect intrusions. Performance of this new approach is measured using the KDDCUP99 dataset and is shown to have high detection and low false alarm rates.
TL;DR: A novel approximate entropy based PseAA composition is proposed to represent apoptosis protein sequences and the results obtained by Jackknife test are quite encouraging, indicating that the proposed method might become a potentially useful tool for protein function, or at least can play a complimentary role to the existing methods in the relevant areas.
Abstract: It is crucial to develop powerful tools to predict apoptosis protein locations for rapidly increasing gap between the number of known structural proteins and the number of known sequences in protein databank. In this study, based on the concept of pseudo amino acid (PseAA) composition originally introduced by Chou, a novel approximate entropy (ApEn) based PseAA composition is proposed to represent apoptosis protein sequences. An ensemble classifier is introduced, of which the basic classifier is the FKNN (fuzzy K-nearest neighbor) one, as prediction engine. Each basic classifier is trained in different dimensions of PseAA composition of protein sequences. The immune genetic algorithm (IGA) is used to search the optimal weight factors in generating the PseAA composition for crucial of weight factors in PseAA composition. The results obtained by Jackknife test are quite encouraging, indicating that the proposed method might become a potentially useful tool for protein function, or at least can play a complimentary role to the existing methods in the relevant areas.
TL;DR: It is proved that under certain conditions, the threshold of an image calculated with the algorithm of Huang and Wang is always the same, that is, it remains invariant.
Abstract: In this paper, we present the definition of restricted dissimilarity function. This definition arises from the concepts of dissimilarity and equivalence function. We analyze the relation there is between restricted dissimilarity functions, restricted equivalence functions (see [Bustince, H., Barrenechea, E., Pagola, M., 2006. Restricted equivalence functions. Fuzzy Sets Syst. 157, 2333-2346]) and normal E"N-functions. We present characterization theorems from implication operators and automorphisms. Next, by aggregating restricted dissimilarity functions in a special way, we construct distance measures of Liu, proximity measures of Fan et al. and fuzzy entropies. We also study diverse interrelations between the above-mentioned concepts. These interrelations enable us to prove that under certain conditions, the threshold of an image calculated with the algorithm of Huang and Wang [Huang, L.K., Wang, M.J., 1995. Image thresholding by minimizing the measure of fuzziness. Pattern Recognit. 28 (1), 41-51], with the methods of Forero [Forero, M.G., 2003. Fuzzy thresholding and histogram analysis. In: Nachtegael, M., Van der Weken, D., Van de Ville, D., Kerre, E.E. (Eds.), Fuzzy Filters for Image Processing. Springer, pp. 129-152] or with the algorithms developed in [Bustince, H., Barrenechea, E., Pagola, M., 2007. Image thresholding using restricted equivalence functions and maximizing the measures of similarity. Fuzzy Sets Syst. 158, 496-516] is always the same, that is, it remains invariant.
TL;DR: Experimental results show that the proposed Log-Gabor filtering method can effectively improve the fingerprint image quality and promote the reliability of fingerprint identification.
Abstract: The performance of automatic fingerprint identification system relies heavily on the quality of fingerprint images. Fingerprint enhancement is essential to ensure the robustness of fingerprint identification with respect to the image quality. Gabor filtering is the most popular method in fingerprint enhancement. To overcome the limitations of traditional Gabor filter and promote fingerprint enhancement performance, the Log-Gabor filter is introduced in this paper. The design method and implementation scheme of Log-Gabor filter in fingerprint enhancement are described in detail. The enhancement performance is assessed on standard fingerprint databases. Experimental results show that the proposed Log-Gabor filtering method can effectively improve the fingerprint image quality and promote the reliability of fingerprint identification.
TL;DR: This paper applies pattern recognition techniques for fast detection of packed executables to efficiently and accurately distinguish between packed and non-packed executables, so that only executables detected as packed will be sent to an universal unpacker, thus saving a significant amount of processing time.
Abstract: Executable packing is the most common technique used by computer virus writers to obfuscate malicious code and evade detection by anti-virus software. Universal unpackers have been proposed that can detect and extract encrypted code from packed executables, therefore potentially revealing hidden viruses that can then be detected by traditional signature-based anti-virus software. However, universal unpackers are computationally expensive and scanning large collections of executables looking for virus infections may take several hours or even days. In this paper we apply pattern recognition techniques for fast detection of packed executables. The objective is to efficiently and accurately distinguish between packed and non-packed executables, so that only executables detected as packed will be sent to an universal unpacker, thus saving a significant amount of processing time. We show that our system achieves very high detection accuracy of packed executables with a low average processing time.
TL;DR: The experiments conducted with 36 real-world data sets available from the UCI repository demonstrate that RotBoost can generate ensemble classifiers with significantly lower prediction error than either Rotation Forest or AdaBoost more often than the reverse.
Abstract: This paper presents a novel ensemble classifier generation technique RotBoost, which is constructed by combining Rotation Forest and AdaBoost. The experiments conducted with 36 real-world data sets available from the UCI repository, among which a classification tree is adopted as the base learning algorithm, demonstrate that RotBoost can generate ensemble classifiers with significantly lower prediction error than either Rotation Forest or AdaBoost more often than the reverse. Meanwhile, RotBoost is found to perform much better than Bagging and MultiBoost. Through employing the bias and variance decompositions of error to gain more insight of the considered classification methods, RotBoost is seen to simultaneously reduce the bias and variance terms of a single tree and the decrement achieved by it is much greater than that done by the other ensemble methods, which leads RotBoost to perform best among the considered classification procedures. Furthermore, RotBoost has a potential advantage over AdaBoost of suiting parallel execution.
TL;DR: A novel approach to multi-texture image segmentation based on the formation of an effective texture feature vector that overcomes the weakness of the single frequency output component of the filter.
Abstract: We present a novel approach to multi-texture image segmentation based on the formation of an effective texture feature vector. Texture sub-features are derived from the output of an optimized Gabor filter. The filter's parameters are selected by an immune genetic algorithm, which aims at maximizing the discrimination between the multi-textured regions. Next the texture features are integrated with a local binary pattern, to form an effective texture descriptor with low computational cost, which overcomes the weakness of the single frequency output component of the filter. Finally, a K-nearest neighbor classifier is used to effect the multi-texture segmentation. The integration of the optimum Gabor filter and local binary pattern methods provide a novel solution to the task. Experimental results demonstrate the effectiveness of the proposed approach.
TL;DR: A multiplicative updating algorithm which iteratively minimizes the @a-divergence between X and AS is derived and it is shown that the same algorithm can be also derived using Karush-Kuhn-Tucker conditions as well as the projected gradient.
Abstract: Non-negative matrix factorization (NMF) is a popular technique for pattern recognition, data analysis, and dimensionality reduction, the goal of which is to decompose non-negative data matrix X into a product of basis matrix A and encoding variable matrix S with both A and S allowed to have only non-negative elements. In this paper, we consider Amari's @a-divergence as a discrepancy measure and rigorously derive a multiplicative updating algorithm (proposed in our recent work) which iteratively minimizes the @a-divergence between X and AS. We analyze and prove the monotonic convergence of the algorithm using auxiliary functions. In addition, we show that the same algorithm can be also derived using Karush-Kuhn-Tucker (KKT) conditions as well as the projected gradient. We provide two empirical study for image denoising and EEG classification, showing the interesting and useful behavior of the algorithm in cases where different values of @a (@a=0.5,1,2) are used.
TL;DR: Experimental results indicate that the proposed edge detection improvement approach is efficient on compensating broken edges and more efficient than the traditional ACO approach in computation reduction.
Abstract: Edge detection is a technique for marking sharp intensity changes, and is important in further analyzing image content. However, traditional edge detection approaches always result in broken pieces, possibly the loss of some important edges. This study presents an ant colony optimization based mechanism to compensate broken edges. The proposed procedure adopts four moving policies to reduce the computation load. Remainders of pheromone as compensable edges are then acquired after finite iterations. Experimental results indicate that the proposed edge detection improvement approach is efficient on compensating broken edges and more efficient than the traditional ACO approach in computation reduction.
TL;DR: It is concluded that the most suitable algorithms for achieving illumination compensation and normalization in eigenspace-based face recognition are SQI and the modified LBP transform.
Abstract: The aim of this work is to investigate illumination compensation and normalization in eigenspace-based face recognition by carrying out an independent comparative study among several pre-processing algorithms. This research is motivated by the lack of direct and detailed comparisons of those algorithms in equal working conditions. The results of this comparative study intend to be a guide for the developers of face recognitions systems. The study focuses on algorithms with the following properties: (i) general purpose, (ii) no modeling steps or training images required, (iii) simplicity, (iv) high speed, and (v) high performance in terms of recognition rates. Thus, herein five different algorithms are compared, by using them as a pre-processing stage in 16 different eigenspace-based face recognition systems. The comparative study is carried out in a face identification scenario using a large amount of images from the PIE, Yale B and Notre Dame face databases. As a result of this study we concluded that the most suitable algorithms for achieving illumination compensation and normalization in eigenspace-based face recognition are SQI and the modified LBP transform.
TL;DR: A new heuristic to combine any type of one-class models for solving the multi-class classification problem with outlier rejection is proposed, which normalizes the average model output per class, instead of the more common non-linear transformation of the distances.
Abstract: In many classification problems objects should be rejected when the confidence in their classification is too low. An example is a face recognition problem where the faces of a selected group of people have to be classified, but where all other faces and non-faces should be rejected. These problems are typically solved by estimating the class densities and assigning an object to the class with the highest posterior probability. The total probability density is thresholded to detect the outliers. Unfortunately, this procedure does not easily allow for class-dependent thresholds, or for class models that are not based on probability densities but on distances. In this paper we propose a new heuristic to combine any type of one-class models for solving the multi-class classification problem with outlier rejection. It normalizes the average model output per class, instead of the more common non-linear transformation of the distances. It creates the possibility to adjust the rejection threshold per class, and also to combine class models that are not (all) based on probability densities and to add class models without affecting the boundaries of existing models. Experiments show that for several classification problems using class-specific models significantly improves the performance.
TL;DR: It is shown through a substantial data-based study of classification accuracy that MIA exhibits consistently good performance across a broad range of data types and of sources and amounts of missingness.
Abstract: We propose a simple and effective method for dealing with missing data in decision trees used for classification. We call this approach ''missingness incorporated in attributes'' (MIA). It is very closely related to the technique of treating ''missing'' as a category in its own right, generalizing it for use with continuous as well as categorical variables. We show through a substantial data-based study of classification accuracy that MIA exhibits consistently good performance across a broad range of data types and of sources and amounts of missingness. It is competitive with the best of the rest (particularly, a multiple imputation EM algorithm method; EMMI) while being conceptually and computationally simpler. A simple combination of MIA and EMMI is slower but even more accurate.
TL;DR: The combination of Gabor features with nearest neighbor or SVM classifier shows promising results; i.e., over 98% for bi-script and tri-script cases and above 89% for the eleven-script scenario.
Abstract: We report an algorithm to identify the script of each word in a document image. We start with a bi-script scenario which is later extended to tri-script and then to eleven-script scenarios. A database of 20,000 words of different font styles and sizes has been collected and used for each script. Effectiveness of Gabor and discrete cosine transform (DCT) features has been independently evaluated using nearest neighbor, linear discriminant and support vector machines (SVM) classifiers. The combination of Gabor features with nearest neighbor or SVM classifier shows promising results; i.e., over 98% for bi-script and tri-script cases and above 89% for the eleven-script scenario.
TL;DR: The proposed Wasserstein-based distance generalizes a wide set of distances proposed for interval data by different approaches or in different contexts of analysis, and shows its interesting properties in the context of clustering techniques.
Abstract: Interval data allow statistical units to be described by means of intervals of values, whereas their representation by means of a single value appears to be too reductive or inconsistent. In the present paper, we present a Wasserstein-based distance for interval data, and we show its interesting properties in the context of clustering techniques. We show that the proposed distance generalizes a wide set of distances proposed for interval data by different approaches or in different contexts of analysis. An application on real data is performed to illustrate the impact of using different metrics and the proposed one using a dynamic clustering algorithm.
TL;DR: A novel video sequence matching method based on temporal ordinal measurements that gives a global and local description of temporal variation and a quantitative method to measure the robustness and discriminability attributes of the matching methods.
Abstract: This paper proposes a novel video sequence matching method based on temporal ordinal measurements. Each frame is divided into a grid and corresponding grids along a time series are sorted in an ordinal ranking sequence, which gives a global and local description of temporal variation. A video sequence matching means not only finding which video a query belongs to, but also a precise temporal localization. Robustness and discriminability are two important issues of video sequence matching. A quantitative method is also presented to measure the robustness and discriminability attributes of the matching methods. Experiments are conducted on a BBC open news archive with a comparison of several methods.
TL;DR: A natural generalization of the Wilcoxon-Mann-Whitney statistic, which now corresponds to the volume under an r-dimensional surface (VUS) for r ordered categories and differs from extensions recently proposed for multi-class classification.
Abstract: Nowadays the area under the receiver operating characteristics (ROC) curve, which corresponds to the Wilcoxon-Mann-Whitney test statistic, is increasingly used as a performance measure for binary classification systems. In this article we present a natural generalization of this concept for more than two ordered categories, a setting known as ordinal regression. Our extension of the Wilcoxon-Mann-Whitney statistic now corresponds to the volume under an r-dimensional surface (VUS) for r ordered categories and differs from extensions recently proposed for multi-class classification. VUS rather evaluates the ranking returned by an ordinal regression model instead of measuring the error rate, a way of thinking which has especially advantages with skew class or cost distributions. We give theoretical and experimental evidence of the advantages and different behavior of VUS compared to error rate, mean absolute error and other ranking-based performance measures for ordinal regression. The results demonstrate that the models produced by ordinal regression algorithms minimizing the error rate or a preference learning based loss, not necessarily impose a good ranking on the data.