Ecologists can assess the health of flooded habitats or wetlands by studying the variations in the populations of bioindicators such as anurans (i.e., frogs and toads). To monitor anuran populations, ecologists manually identify anuran species from audio recordings. This identification task can be significantly streamlined by the availability of an automated method for anuran identification. Previous promising frog-call identification methods have relied on the extraction of pre-designed features from audio spectrograms such as Mel Coefficients and other filter responses. Instead of using pre-designed features, we propose to allow a deep-learning algorithm to find the features that are most important for classification. In work reported in this paper, we used two deep-learning methods that apply convolutional neural networks (CNN) to anuran classification. Transfer learning was also used. The CNN methods was tested on our dataset of 15 frog species, and produced a classification accuracy up to 77%.

Ecologists can assess the health of wetlands by monitoring populations of animals such as Anurans (i.e., frogs and toads), which are sensitive to habitat changes. But, surveying anurans requires trained experts to identify species from the animals' mating calls. This identification task can be streamlined by automation. To this end, we propose an automatic frog-call classification algorithm and a smartphone application that drastically simplify the monitoring of anuran populations. We offer three main contributions. First, we introduce a classification method that has an average accuracy of 86% on a dataset of 736 calls from 48 anuran species from the United States. Our dataset is much larger and diverse than those of previous works on anuran classification. Second, we extract a new type of spectrogram feature that avoids syllable segmentation and the manual cleaning of the recordings. Our method also works with recordings of variable length. Third, our method uses GPS location and a voting scheme to reliably deal with a large number of species and high levels of noise.

Starting from an object's location in a video frame, tracking-by-detection methods find the location of that object in a subsequent video frame. The tracker's detection step may produce multiple false positives during short-term occlusions, which can result in loss of track. We propose a tracking-by-detection method that is robust to short-term occlusions and false positives. Here, we extend the Struck tracker, which is based on structured SVM, to output bounding boxes at multiple scales. In addition to predicting scale changes, we use the Robust Kalman filter to decrease false-positive detections and to increase the tracker resilience to short-time occlusions. Here, we develop a special strategy for the tracker's update step, which is designed to decrease over fitting and to allow for the tracker to recover from loss of track. We thoroughly evaluate our method on a publicly available video dataset and show that it outperforms the state-of-the-art.

In this paper, we address the problem of visual tracking in videos without using a pre-learned model of the object. This type of model-free tracking is a hard problem because of limited information about the object, abrupt object motion, and shape deformation. We propose to integrate an object-agnostic prior, called objectness, which is designed to measure the likelihood of a given location to contain an object of any type, into structured tracking framework. Our objectness prior is based on image segmentation and edges; thus, it does not require training data. By extending a structured tracker with the prior, we introduce a new tracker which we call ObjStruck. We extensively evaluate our tracker on publicly available datasets and show that objectness prior improves tracking accuracy.

We propose a method to recognize pollen grains using a two-stage classifier. First, texture classification categorizes the pollen grains into sub-groups. Then, a final classification of individual pollen types is done by segmenting the image int multiple layers of regions for each pollen image. The main novelty in our method is threefold: (1) Adopting two successive classification stages. (2) Combining hierarchical clustering and SVM algorithms to merge similar pollen types into sub-groups. (3) Adopting a layering approach prior to performing feature extraction. The combination of these aspects gives excellent results. We evaluated our method using 1,063 light-microscopy images of pollen grains from 30 species. The results show that: (1) the layering technique increases the classification rate by almost almost 7% over using the same features directly. (2) adopting two classification stage increases the classification rate by 6%. (3) the proposed system outperformed traditional techniques.

We present a novel technique for pollen identification from sets of multifocal image sequences obtained from optical microscopy. Our algorithm analyzes the visual texture of pollen grains for each focal image, and performs identification using a fast sequence-matching algorithm. Although we develop a pollen-recognition protocol, the method is applicable to other microscopy object-recognition tasks. The proposed method requires little manual interaction, and does not rely on specialized imaging procedures such as florescence and deconvolution. We test our method on images of tropical fossil pollen.

Football (aka Soccer) is the most popular sport in the world. The popularity of the sport leads to several stories (some perhaps anecdotal) about supporters behaviors and to the emergence of rivalries such as the famous Barcelona-Real Madrid (in Spain). Little however has been done to characterize/profile online users' behaviors as football supporters and use them as an aggregate measure to club characterization. Today, the availability of data enable us to understand at a much greater scale if rivalries exist and if there are signatures that can be used to characterize supporting behavior. In this paper we use techniques from Data Science to characterize football supporters according to their activity on Twitter and to characterize clubs according to the behavior of their supporters. We show that it is possible to: (i) rank football clubs by their popularity and fans' dislike, (ii) identify the rivalries that exist between clubs and their supporters, and (iii) find specific signatures that repeat themselves across different clubs and in different countries. The results are evaluated on a large dataset of tweets relevant to major football leagues in Brazil and in the United Kingdom.

We propose a method for classifying actions involving people interacting with objects. Our method combines motion and appearance information into a unified framework. Here, we explore the video's sparse component as provided by robust principal-component analysis for the extraction of motion information in the form of trajectories. While we use motion as the main clue for classification, we also incorporate implicit object information into the classification process. Here, object information is represented by the probability of the object with which the person is interacting. These probabilities are learned using probabilistic Latent Semantic Analysis (pLSA). We test our classification method on a publicly available dataset, and provide a comparison with some related work. Classification results obtained by our method are promising.

Image moments have been widely used for designing robust shape descriptors that are invariant to rigid transformations. In this work, we address the problem of estimating non-rigid deformation fields based on image moment variations. By using a single family of polynomials to both parameterize the deformation field and to define image moments, we can represent image moments variation as a system of quadratic functions, and solve for the deformation parameters. As a result, we can recover the deformation field between two images without solving the correspondence problem. Additionally, our method is highly robust to image noise. The method was tested on both synthetically deformed MPEG-7 shapes and cardiac MRI sequences.

A number of action recognition methods make use of spatio-temporal features. These features often consist of local spatio-temporal descriptors centered at locations provided by an interest point detector. The extracted descriptors will then serve as input to classification algorithms. The correct scale of these descriptors is an essential parameter to be determined. Improved information quality has been achieved from recently developed entropy-based spatio-temporal feature descriptors. In this paper, we present an approach for determining scales of the sub-volumes of interest given the locations of spatio-temporal features. Our method works by measuring the average variations of local motion content calculated on subsequences of motion filter responses. We design a filter-specific data prior that allows to determine the scales of the informative neighborhoods. We demonstrate that features calculated at the scales provided by or method allow for noticeable performance improvements of action recognition algorithms.

We present a novel method for learning human motion models from unsegmented videos. We propose a unified framework that encodes spatio-temporal relationships between descriptive motion parts and the appearance of individual poses. Sparse sets of spatial and spatio-temporal features are used. The method automatically learns static pose models and spatio-temporal motion parts. Neither motion cycles nor human figures need to be segmented for learning. We test the model on a publicly available action dataset and demonstrate that our new method performs well on a number of classification tasks. We also show that classification rates are improved by increasing the number of pose models in the framework.

In this paper, we present a solution to the novel problem of recognizing primitive actor-object interactions from videos. Here, we introduce the concept of actor-object states. Our method is based on the observation that at the moment of physical contact, both the motion and the appearance of actors are constrained by the target object. We propose a probabilistic framework that automatically learns models in such constrained states. We use joint probability distributions to represent both actor and object appearances as well as their intrinsic spatio-temporal configurations. Finally, we demonstrate the applicability of our approach on series of human-object interaction classification experiments.

In this paper we investigate how texture recognition can be achieved through the modal analysis of the pattern of peaks in the spectral density function. We commence from a texture characterisation which is based on the positions of peaks in the power spectrum. Our aim is to use the modal structure of the pattern of peaks to perform texture retrieval from an image data-base. We explore two different approaches to the problem. First, we use a variant of the Shapiro and Brady method to perform recognition by comparing the modal structure of the proximity matrix for peak cluster centres. Second, we perform latent semantic indexing on vectors representing the polar distribution of frequency peaks. We provide and experimental evaluation of these two methods on a data-base of fabric and wrapping paper patterns.

Shows how the local slant and tilt angles of regularly textured curved surfaces can be estimated directly, without the need for iterative numerical optimization. We work in the frequency domain and measure texture distortion using the affine distortion of the pattern of spectral peaks. The key theoretical contribution is to show that the directions of the eigenvectors of the affine distortion matrices can be used to estimate local slant and tilt angles of tangent planes to curved surfaces. In particular, the leading eigenvector points in the tilt direction. Although not as geometrically transparent, the direction of the second eigenvector can be used to estimate the slant direction. The required affine distortion matrices are computed using the correspondences between spectral peaks, established on the basis of their energy ordering. We apply the method to a variety of real-world and synthetic imagery.

Multiple vanishing point detection provides the key to recovering the perspective pose of textured planes. If vanishing points are to be detected from spectral information then there are two computational problems that need to be solved. Firstly, the search of the extended image plane is unbounded, and hence the location of vanishing points at or near infinity is difficult. Secondly, correspondences between local spectra need to be established so that vanishing points can be triangulated. In this paper we offer a way of overcoming these two difficulties. We overcome the problem of unbounded search by mapping the information provided by local spectral moments onto a unit-sphere. According to our representation, the position and direction of each local spectrum maps onto a great circle on the unit-sphere. The need for correspondences is overcome by accumulating the great circle intercepts. Vanishing points occur at local accumulator maxima on the unit-sphere. We experiment with the new shape-from-texture technique on planar textures in buildings.

Spectral analysis provides a powerful means of estimating the perspective pose of texture planes. Unfortunately one of the problems that restricts the utility of the method is the need to set the size of the spectral window. For texture planes viewed under extreme perspective distortion, the spectral frequency density may vary rapidly across the image plane. If the size of the window is mismatched to the underlying texture distribution, then the estimated frequency spectrum may become severely defocused. This in turn limits the accuracy of perspective pose estimation. The aim in this paper is to describe an adaptive method for setting the size of the spectral window. We provide an analysis which shows that there is a window size that minimises the degree of defocusing.

Multiple vanishing point detection provides the key to recovering the perspective pose of textured planes. If vanishing points are to be detected from spectral information then there are two computational problems that need to be solved. Firstly, the search of the extended image plane is unbounded, and hence the location of vanishing points at or near infinity is difficult. Secondly, correspondences between local spectra need to be established so that vanishing points can be triangulated. We offer a way of overcoming these two difficulties. We overcome the problem of unbounded search by mapping the information provided by local spectral moments onto a unit-sphere. According to our representation, the position and direction of each local spectrum maps onto a great circle on the unit-sphere. The need for correspondences is overcome by accumulating the great circle intercepts. Vanishing points occur at local accumulator maxima on the unit-sphere. We experiment with the new shape-from-texture technique on planar textures in buildings.

Spectral analysis provides a powerful means of estimating the perspective pose of texture planes. However, one of the problems that restricts the utility of the method is the need to set the size of the spectral window. This paper describes an adaptive method for setting the size of the spectral window. We provide an analysis which shows that there is a window size that minimises the degree of defocusing of the power spectrum. The minimum is located through an analysis of the spectral covariance matrix. We experiment with the new method on both synthetic and real world imagery. This demonstrates that the method provides accurate pose angle estimates, even when the slant angle is large. We also provide a comparison of the accuracy of perspective pose estimation that results both from our adaptive scale method and with one of fixed scale.

Presents an approach to the recovery of dense orientation estimates for curved textured surfaces. We show how the affine distance between pairs of power spectra can be used to estimate the slant and tilt angles of local tangent planes to the textured surface. The main practical benefit furnished by our analysis is that it allows us to estimate the orientation angles in closed form without recourse to numerical optimisation. Based on these theoretical properties we present an algorithm for the analysis of regularly textured curved surfaces. We apply the method to a variety of real-world and synthetic imagery. We show that the new shape-from-texture method can reliably estimate surface topography.

Multiple vanishing point detection provides the key to recovering the perspective pose of textured planes. If vanishing points are to be detected from spectral information then there are two computational problems that need to be solved. Firstly, the searchofthe extended image plane is unbounded, and hence the location of vanishing points at or near infinity is difficult. Secondly, correspondences between local spectra need to be established so that vanishing points can be triangulated. In this paper we offer a way of overcoming these two difficulties. We overcome the problem of unbounded search by mapping the information provided by local spectral moments onto a unitsphere.

Vanishing point locations play a critical role in determining the perspective orientation of textured planes. However, if only a single vanishing point is available then the problem is undetermined. The tilt direction must be computed using supplementary information such as the texture gradient. In this payer we describe a method for multiple vanishing point and, hence complete perspective pose estimation, detection which obviates the need to compute the texture gradient. The method is based on local spectral analysis. It exploits the fact that spectral orientation is uniform along lines that radiate from vanishing points on the image place. We demonstrate that the method provides accurate pose angle estimates, even when the slant angle is large.

This paper describes a novel method for recovering the perspective geometry of textured surfaces from local spectral moments. It commences by computing the Fourier power spectrum at each of a series of local neighbourhoods.