Kernelized structural classification for 3D dogs body parts detection

Despite pattern recognition methods for human behavioral analysis has flourished in the last decade, animal behavioral analysis has been almost neglected. Those few approaches are mostly focused on preserving livestock economic value while attention on the welfare of companion animals, like dogs, is now emerging as a social need. In this work, following the analogy with human behavior recognition, we propose a system for recognizing body parts of dogs kept in pens. We decide to adopt both 2D and 3D features in order to obtain a rich description of the dog model. Images are acquired using the Microsoft Kinect to capture the depth map images of the dog. Upon depth maps a Structural Support Vector Machine (SSVM) is employed to identify the body parts using both 3D features and 2D images. The proposal relies on a kernelized discriminative structural classificator specifically tailored for dogs independently from the size and breed. The classification is performed in an online fashion using the LaRank optimization technique to obtaining real time performances. Promising results have emerged during the experimental evaluation carried out at a dog shelter, managed by IZSAM, in Teramo, Italy.

Recurrent Convolutional Network for Video-based Person Re-Identification

In this paper we propose a novel recurrent neural networkarchitecture for video-based person re-identification.Given the video sequence of a person, features are extracted from each frame using a convolutional neural network that incorporates a recurrent final layer, which allows information to flow between time-steps. The features from all time steps are then combined using temporal pooling to give an overall appearance feature for the complete sequence. The convolutional network, recurrent layer, and temporal pooling layer, are jointly trained to act as a feature extractor for video-based re-identification using a Siamese network architecture.Our approach makes use of colour and optical flow information in order to capture appearance and motion information which is useful for video re-identification. Experiments are conduced on the iLIDS-VID and PRID-2011 datasets to show that this approach outperforms existing methods of video-based re-identification.

https://github.com/niallmcl/Recurrent-Convolutional-Video-ReID
Project Source Code

k-fold Subsampling based Sequential Backward Feature Elimination

We present a new wrapper feature selection algorithm for human detection. This algorithm is a hybrid featureselection approach combining the benefits of filter and wrapper methods. It allows the selection of an optimalfeature vector that well represents the shapes of the subjects in the images. In detail, the proposed featureselection algorithm adopts the k-fold subsampling and sequential backward elimination approach, while thestandard linear support vector machine (SVM) is used as the classifier for human detection. We apply theproposed algorithm to the publicly accessible INRIA and ETH pedestrian full image datasets with the PASCALVOC evaluation criteria. Compared to other state of the arts algorithms, our feature selection based approachcan improve the detection speed of the SVM classifier by over 50% with up to 2% better detection accuracy.Our algorithm also outperforms the equivalent systems introduced in the deformable part model approach witharound 9% improvement in the detection accuracy

Automatic information and safety systems for driving assistance

O objeto principal desta tese é o estudo de algoritmos de processamento e representação automáticos de dados, em particular de informação obtida por sensores montados a bordo de veículos (2D e 3D), com aplicação em contexto de sistemas de apoio à condução. O trabalho foca alguns dos problemas que, quer os sistemas de condução automática (AD), quer os sistemas avançados de apoio à condução (ADAS), enfrentam hoje em dia. O documento é composto por duas partes. A primeira descreve o projeto, construção e desenvolvimento de três protótipos robóticos, incluindo pormenores associados aos sensores montados a bordo dos robôs, algoritmos e arquitecturas de software. Estes robôs foram utilizados como plataformas de ensaios para testar e validar as técnicas propostas. Para além disso, participaram em várias competições de condução autónoma tendo obtido muito bons resultados. A segunda parte deste documento apresenta vários algoritmos empregues na geração de representações intermédias de dados sensoriais. Estes podem ser utilizados para melhorar técnicas já existentes de reconhecimento de padrões, deteção ou navegação, e por este meio contribuir para futuras aplicações no âmbito dos AD ou ADAS. Dado que os veículos autónomos contêm uma grande quantidade de sensores de diferentes naturezas, representações intermédias são particularmente adequadas, pois podem lidar com problemas relacionados com as diversas naturezas dos dados (2D, 3D, fotométrica, etc.), com o carácter assíncrono dos dados (multiplos sensores a enviar dados a diferentes frequências), ou com o alinhamento dos dados (problemas de calibração, diferentes sensores a disponibilizar diferentes medições para um mesmo objeto). Neste âmbito, são propostas novas técnicas para a computação de uma representação multi-câmara multi-modal de transformação de perspectiva inversa, para a execução de correcção de côr entre imagens de forma a obter mosaicos de qualidade, ou para a geração de uma representação de cena baseada em primitivas poligonais, capaz de lidar com grandes quantidades de dados 3D e 2D, tendo inclusivamente a capacidade de refinar a representação à medida que novos dados sensoriais são recebidos.

Classification of incomplete feature vectors by radial basis function networks

The paper describes the use of radial basis function neural networks with Gaussian basis functions to classify incomplete feature vectors. The method uses the fact that any marginal distribution of a Gaussian distribution can be determined from the mean vector and covariance matrix of the joint distribution.

Multi-scale cortical keypoint representation for attention and object detection

Keypoints (junctions) provide important information for focus-of-attention (FoA) and object categorization/recognition. In this paper we analyze the multi-scale keypoint representation, obtained by applying a linear and quasi-continuous scaling to an optimized model of cortical end-stopped cells, in order to study its importance and possibilities for developing a visual, cortical architecture.We show that keypoints, especially those which are stable over larger scale intervals, can provide a hierarchically structured saliency map for FoA and object recognition. In addition, the application of non-classical receptive field inhibition to keypoint detection allows to distinguish contour keypoints from texture (surface) keypoints.

Face normalization using multi-scale cortical keypoints

Empirical studies concerning face recognition suggest that faces may be stored in memory by a few canonical representations. Models of visual perception are based on image representations in cortical area V1 and beyond, which contain many cell layers for feature extractions. Simple, complex and end-stopped cells tuned to different spatial frequencies (scales) and/or orientations provide input for line, edge and keypoint detection. This yields a rich, multi-scale object representation that can be stored in memory in order to identify objects. The multi-scale, keypoint-based saliency maps for Focus-of-Attention can be explored to obtain face detection and normalization, after which face recognition can be achieved using the line/edge representation. In this paper, we focus only on face normalization, showing that multi-scale keypoints can be used to construct canonical representations of faces in memory.

Fast segmentation of 3D data using an octree

The algorithm developed uses an octree pyramid in which noise is reduced at the expense of the spatial resolution. At a certain level an unsupervised clustering without spatial connectivity constraints is applied. After the classification, isolated voxels and insignificant regions are removed by assigning them to their neighbours. The spatial resolution is then increased by the downprojection of the regions, level by level. At each level the uncertainty of the boundary voxels is minimised by a dynamic selection and classification of these, using an adaptive 3D filtering. The algorithm is tested using different data sets, including NMR data.

Disparity energy model with keypoint disparity validation

A biological disparity energy model can estimate local depth information by using a population of V1 complex cells. Instead of applying an analytical model which explicitly involves cell parameters like spatial frequency, orientation, binocular phase and position difference, we developed a model which only involves the cells’ responses, such that disparity can be extracted from a population code, using only a set of previously trained cells with random-dot stereograms of uniform disparity. Despite good results in smooth regions, the model needs complementary processing, notably at depth transitions. We therefore introduce a new model to extract disparity at keypoints such as edge junctions, line endings and points with large curvature. Responses of end-stopped cells serve to detect keypoints, and those of simple cells are used to detect orientations of their underlying line and edge structures. Annotated keypoints are then used in the leftright matching process, with a hierarchical, multi-scale tree structure and a saliency map to segregate disparity. By combining both models we can (re)define depth transitions and regions where the disparity energy model is less accurate.

Joint inversion of magnetotelluric and seismic data for crustal characterization

Tese de doutoramento, Ciências Geofísicas e da Geoinformação (Geofisíca), Universidade de Lisboa, Faculdade de Ciências, 2014

Adaptive-threshold region merging via path scanning

Region merging algorithms commonly produce results that are seen to be far below the current commonly accepted state-of-the-art image segmentation techniques. The main challenging problem is the selection of an appropriate and computationally efficient method to control resolution and region homogeneity. In this paper we present a region merging algorithm that includes a semi-greedy criterion and an adaptive threshold to control segmentation resolution. In addition we present a new relative performance indicator that compares algorithm performance across many metrics against the results from human segmentation. Qualitative (visual) comparison demonstrates that our method produces results that outperform existing leading techniques.

Role on serine protease inhibitors in the pathogenisis of Steinernema carpocapsae

Tese de Doutoramento, Biologia (Biologia Celular e Molecular), 18 de Novembro de 2013, Universidade dos Açores.

Clustering de dados biomédicos com algoritmos baseados em critérios entrópicos

A procura de padrões nos dados de modo a formar grupos é conhecida como aglomeração de dados ou clustering, sendo uma das tarefas mais realizadas em mineração de dados e reconhecimento de padrões. Nesta dissertação é abordado o conceito de entropia e são usados algoritmos com critérios entrópicos para fazer clustering em dados biomédicos. O uso da entropia para efetuar clustering é relativamente recente e surge numa tentativa da utilização da capacidade que a entropia possui de extrair da distribuição dos dados informação de ordem superior, para usá-la como o critério na formação de grupos (clusters) ou então para complementar/melhorar algoritmos existentes, numa busca de obtenção de melhores resultados. Alguns trabalhos envolvendo o uso de algoritmos baseados em critérios entrópicos demonstraram resultados positivos na análise de dados reais. Neste trabalho, exploraram-se alguns algoritmos baseados em critérios entrópicos e a sua aplicabilidade a dados biomédicos, numa tentativa de avaliar a adequação destes algoritmos a este tipo de dados. Os resultados dos algoritmos testados são comparados com os obtidos por outros algoritmos mais “convencionais" como o k-médias, os algoritmos de spectral clustering e um algoritmo baseado em densidade.