Catálogo de publicaciones - libros

We present a novel semi-local image descriptor which encodes multiple edges corresponding to the image structure boundaries around an interest point. The proposed method addresses the problem of poor edge detection through a robust, scale and orientation invariant, descriptor distance. In addition, a clustering of descriptors capable of extracting distinctive shapes from a set of descriptors is described. The proposed techniques are applied to the description of bone shapes in medical X-ray images and the experimental results are presented.

This paper proposes an unsupervised learning technique for object recognition from an unlabelled and unordered set of training images. It enables the robust recognition of complex 3D objects in cluttered scenes, under scale changes and partial occlusion. The technique uses a matching based on the consistency of two different descriptors characterising the appearance and shape of local features. The variation of each local feature with viewing direction is modeled by a multi-view feature model. These multi-view feature models can be matched directly to the features found in a test image. This avoids a matching to all training views as necessary for approaches based on canonical views.

The proposed approach is tested with real world objects and compared to a supervised approach using features characterised by SIFT descriptors (Scale Invariant Feature Transform). These experiments show that the performance of our unsupervised technique is equal to that of a supervised SIFT object recognition approach.

This work presents a new approach to analysis of shapes represented by finite set of landmarks, that generalizes the notion of Procrustes distance - an invariant metric under translation, scaling, and rotation. In many shape classification tasks there is a large variability in certain landmarks due to intra-class and/or inter-class variations. Such variations cause poor shape alignment needed for Procrustes distance computation, and lead to poor classification performance. We apply a general framework to the task of supervised classification of shapes that naturally deals with landmark distributions exhibiting large intra class or inter-class variabilty. The incorporation of Procrustes metric and of a learnt general quadratic distance inspired by Fisher linear discriminant objective function, produces a generalized Procrustes distance. The learnt distance retains the invariance properties and emphasizes the discriminative shape features. In addition, we show how the learnt metric can be useful for kernel machines design and demonstrate a performance enhancement accomplished by the learnt distances on a variety of classification tasks of organismal forms datasets.

Recent solutions to object classification have focused on the decomposition of objects into representative parts. However, the vast majority of these methods are based on single visual cue measurements. Psychophysical evidence suggests that humans use multiple visual cues to accomplish recognition. In this paper, we address the problem of integrating multiple visual information for object recognition. Our contribution in this paper is twofold. First, we describe a new probabilistic integration model of multiple visual cues at different spatial locations across the image. Secondly, we use the cue integration framework to classify images of objects by combining two-dimensional and three-dimensional visual cues. Classification results obtained using the method are promising.

We propose a modification of RANSAC that performs guided search of the sample space. The sampling is applicable to any of the sample consensus methods, such as MAPSAC or MLESAC. We give simulation results which show that the new method can reduce the number of required iterations to find a good model by orders of magnitude.

An efficient algorithm to detect, correlate, and track features in a scene was implemented on an FPGA in order to obtain real-time performance. The algorithm implemented was a Harris Feature Detector combined with a correlator based on a priority queue of feature strengths that considered minimum distances between features. The remaining processing of frame to frame movement is completed in software to determine an affine homography including translation, rotation, and scaling. A RANSAC method is used to remove mismatched features and increase accuracy. This implementation was designed specifically for use as an onboard vision solution in determining movement of small unmanned air vehicles that have size, weight, and power limitations.

In this paper we investigate the quality of 3D-2D pose estimates using hand labeled line and point correspondences. We select point correspondences from junctions in the image, allowing to construct a meaningful interpretation about how the junction is formed, as proposed in e.g. [1], [2], [3]. We make us of this information referred as the semantic interpretation, to identify the different types of junctions (i.e. L-junctions and T-junctions). T-junctions often denote occluding contour, and thus do not designate a point in space. We show that the semantic interpretations is useful for the removal of these T-junction from correspondence sets, since they have a negative effect on motion estimates. Furthermore, we demonstrate the possibility to derive additional line correspondences from junctions using the semantic interpretation, providing more constraints and thereby more robust estimates.

Shape from texture has received much attention in the past few decades. We propose a computationally efficient method to extract the 3D shape of developable surfaces from the spectral variations of a visual texture. Under the assumption of homogeneity, the texture is represented by the novel method of identifying ridges of its Fourier transform. Local spatial frequencies are then computed using a minimal set of selected Gabor filters. In both orthographic and perspective projection cases, new geometric equations are presented to compute the shape of developable surfaces from frequencies. The results are validated with semi-synthetic and real pictures.

Image-based full body 3D reconstruction for tele-immersive applications generates large amount of data points, which have to be sent through the network in real-time. In this paper we introduce a skeleton-based compression method using motion estimation where kinematic parameters of the human body are extracted from the point cloud data in each frame. First we address the issues regarding the data capturing and transfer to a remote site for the tele-immersive collaboration. We compare the results of the existing compression methods and the proposed skeleton-based compression technique. We examine robustness and efficiency of the algorithm through experimental results with our multi-camera tele-immersion system. The proposed skeleton-based method provides high and flexible compression ratios (from 50:1 to 5000:1) with reasonable reconstruction quality (peak signal-to-noise ratio from 28 to 31 dB).

In this paper we present the utilization of advanced programming techniques on current graphics hardware to improve the performance of remote rendering for interactive applications. We give an overview of existing systems in remote rendering and focus on some general bottlenecks of remote visualization. Afterwards we describe current developments in graphics hardware and software and outline how they can be used to increase the performance of remote graphics systems. Finally we present some results and benchmarks to confirm the validity of our work.