Abstract: Document Visual Question Answering (DocVQA) refers to the task of answering questions from document images. Existing work on DocVQA only considers single-page documents. However, in real scenarios documents are mostly composed of multiple pages that should be processed altogether. In this work we extend DocVQA to the multi-page scenario. For that, we first create a new dataset, MP-DocVQA, where questions are posed over multi-page documents instead of single pages. Second, we propose a new hierarchical method, Hi-VT5, based on the T5 architecture, that overcomes the limitations of current methods to process long multi-page documents. The proposed method is based on a hierarchical transformer architecture where the encoder summarizes the most relevant information of every page and then, the decoder takes this summarized information to generate the final answer. Through extensive experimentation, we demonstrate that our method is able, in a single stage, to answer the questions and provide the page that contains the relevant information to find the answer, which can be used as a kind of explainability measure.

Abstract: Graph-based representations are of broad use and applicability in pattern recognition. They exhibit, however, a major drawback with regards to the processing tools that are available in their domain. Graphembedding into vectorspaces is a growing field among the structural pattern recognition community which aims at providing a feature vector representation for every graph, and thus enables classical statistical learning machinery to be used on graph-based input patterns. In this work, we propose a novel embedding methodology for graphs with continuous nodeattributes and unattributed edges. The approach presented in this paper is based on statistics of the node labels and the edges between them, based on their similarity to a set of representatives. We specifically deal with an important issue of this methodology, namely, the selection of a suitable set of representatives. In an experimental evaluation, we empirically show the advantages of this novel approach in the context of different classification problems using several databases of graphs.

Abstract: Structural pattern recognition approaches offer the most expressive, convenient, powerful but computational expensive representations of underlying relational information. To benefit from mature, less expensive and efficient state-of-the-art machine learning models of statistical pattern recognition they must be mapped to a low-dimensional vector space. Our method of explicit graph embedding bridges the gap between structural and statistical pattern recognition. We extract the topological, structural and attribute information from a graph and encode numeric details by fuzzy histograms and symbolic details by crisp histograms. The histograms are concatenated to achieve a simple and straightforward embedding of graph into a low-dimensional numeric feature vector. Experimentation on standard public graph datasets shows that our method outperforms the state-of-the-art methods of graph embedding for richly attributed graphs.

Abstract: Non-linear dimensionality reduction techniques are affected by two critical aspects: (i) the design of the adjacency graphs, and (ii) the embedding of new test data-the out-of-sample problem. For the first aspect, the proposed solutions, in general, were heuristically driven. For the second aspect, the difficulty resides in finding an accurate mapping that transfers unseen data samples into an existing manifold. Past works addressing these two aspects were heavily parametric in the sense that the optimal performance is only achieved for a suitable parameter choice that should be known in advance. In this paper, we demonstrate that the sparse representation theory not only serves for automatic graph construction as shown in recent works, but also represents an accurate alternative for out-of-sample embedding. Considering for a case study the Laplacian Eigenmaps, we applied our method to the face recognition problem. To evaluate the effectiveness of the proposed out-of-sample embedding, experiments are conducted using the K-nearest neighbor (KNN) and Kernel Support Vector Machines (KSVM) classifiers on six public face datasets. The experimental results show that the proposed model is able to achieve high categorization effectiveness as well as high consistency with non-linear embeddings/manifolds obtained in batch modes.

Abstract: We present ECOC-DRF, a framework where potential functions for Discriminative Random Fields are formulated as an ensemble of classifiers. We introduce the label trick, a technique to express transitions in the pairwise potential as meta-classes. This allows to independently learn any possible transition between labels without assuming any pre-defined model. The Error Correcting Output Codes matrix is used as ensemble framework for the combination of margin classifiers. We apply ECOC-DRF to a large set of classification problems, covering synthetic, natural and medical images for binary and multi-class cases, outperforming state-of-the art in almost all the experiments.

Abstract: Spatial pyramids have been successfully applied to incorporating spatial information into bag-of-words based image representation. However, a major drawback is that it leads to high dimensional image representations. In this paper, we present a novel framework for obtaining compact pyramid representation. First, we investigate the usage of the divisive information theoretic feature clustering (DITC) algorithm in creating a compact pyramid representation. In many cases this method allows us to reduce the size of a high dimensional pyramid representation up to an order of magnitude with little or no loss in accuracy. Furthermore, comparison to clustering based on agglomerative information bottleneck (AIB) shows that our method obtains superior results at significantly lower computational costs. Moreover, we investigate the optimal combination of multiple features in the context of our compact pyramid representation. Finally, experiments show that the method can obtain state-of-the-art results on several challenging data sets.

Abstract: PR4883, PII: S0031-3203(13)00327-0
Two-dimensional shape models have been successfully applied to solve many problems in computer vision, such as object tracking, recognition, and segmentation. Typically, 2D shape models are learned from a discrete set of image landmarks (corresponding to projection of 3D points of an object), after applying Generalized Procustes Analysis (GPA) to remove 2D rigid transformations. However, the
standard GPA process suffers from three main limitations. Firstly, the 2D training samples do not necessarily cover a uniform sampling of all the 3D transformations of an object. This can bias the estimate of the shape model. Secondly, it can be computationally expensive to learn the shape model by sampling 3D transformations. Thirdly, standard GPA methods use only one reference shape, which can might be insufficient to capture large structural variability of some objects.
To address these drawbacks, this paper proposes continuous generalized Procrustes analysis (CGPA).
CGPA uses a continuous formulation that avoids the need to generate 2D projections from all the rigid 3D transformations. It builds an efficient (in space and time) non-biased 2D shape model from a set of 3D model of objects. A major challenge in CGPA is the need to integrate over the space of 3D rotations, especially when the rotations are parameterized with Euler angles. To address this problem, we introduce the use of the Haar measure. Finally, we extended CGPA to incorporate several reference shapes. Experimental results on synthetic and real experiments show the benefits of CGPA over GPA.