Abstract: In Part 1 of the thesis, we hypothesize that salient and non-salient image regions can be estimated to be the regions which are enhanced or assimilated in standard low-level color image representations. We prove this hypothesis by adapting a low-level model of color perception into a saliency estimation model. This model shares the three main steps found in many successful models for predicting attention in a scene: convolution with a set of filters, a center-surround mechanism and spatial pooling to construct a saliency map. For such models, integrating spatial information and justifying the choice of various parameter values remain open problems. Our saliency model inherits a principled selection of parameters as well as an innate spatial pooling mechanism from the perception model on which it is based. This pooling mechanism has been fitted using psychophysical data acquired in color-luminance setting experiments. The proposed model outperforms the state-of-the-art at the task of predicting eye-fixations from two datasets. After demonstrating the effectiveness of our basic saliency model, we introduce an improved image representation, based on geometrical grouplets, that enhances complex low-level visual features such as corners and terminations, and suppresses relatively simpler features such as edges. With this improved image representation, the performance of our saliency model in predicting eye-fixations increases for both datasets.
In Part 2 of the thesis, we investigate the problem of aesthetic visual analysis. While a great deal of research has been conducted on hand-crafting image descriptors for aesthetics, little attention so far has been dedicated to the collection, annotation and distribution of ground truth data. Because image aesthetics is complex and subjective, existing datasets, which have few images and few annotations, have significant limitations. To address these limitations, we have introduced a new large-scale database for conducting Aesthetic Visual Analysis, which we call AVA. AVA contains more than 250,000 images, along with a rich variety of annotations. We investigate how the wealth of data in AVA can be used to tackle the challenge of understanding and assessing visual aesthetics by looking into several problems relevant for aesthetic analysis. We demonstrate that by leveraging the data in AVA, and using generic low-level features such as SIFT and color histograms, we can exceed state-of-the-art performance in aesthetic quality prediction tasks.
Finally, we entertain the hypothesis that low-level visual information in our saliency model can also be used to predict visual aesthetics by capturing local image characteristics such as feature contrast, grouping and isolation, characteristics thought to be related to universal aesthetic laws. We use the weighted center-surround responses that form the basis of our saliency model to create a feature vector that describes aesthetics. We also introduce a novel color space for fine-grained color representation. We then demonstrate that the resultant features achieve state-of-the-art performance on aesthetic quality classification.
As such, a promising contribution of this thesis is to show that several vision experiences – low-level color perception, visual saliency and visual aesthetics estimation – may be successfully modeled using a unified framework. This suggests a similar architecture in area V1 for both color perception and saliency and adds evidence to the hypothesis that visual aesthetics appreciation is driven in part by low-level cues.

Abstract: In the last years, intrinsic image decomposition has gained attention. Most of the state-of-the-art methods are based on the assumption that reflectance changes come along with strong image edges. Recently, user intervention in the recovery problem has proved to be a remarkable source of improvement. In this paper, we propose a novel approach that aims to overcome the shortcomings of pure edge-based methods by introducing strong surface descriptors, such as the color-name descriptor which introduces high-level considerations resembling top-down intervention. We also use a second surface descriptor, termed color-shade, which allows us to include physical considerations derived from the image formation model capturing gradual color surface variations. Both color cues are combined by means of a Markov Random Field. The method is quantitatively tested on the MIT ground truth dataset using different error metrics, achieving state-of-the-art performance.

Abstract: In this article, a novel technique for user’s authentication and verification using gait as a biometric unobtrusive pattern is proposed. The method is based on a two stages pipeline. First, a general activity recognition classifier is personalized for an specific user using a small sample of her/his walking pattern. As a result, the system is much more selective with respect to the new walking pattern. A second stage verifies whether the user is an authorized one or not. This stage is defined as a one-class classification problem. In order to solve this problem, a four-layer architecture is built around the geometric concept of convex hull. This architecture allows to improve robustness to outliers, modeling non-convex shapes, and to take into account temporal coherence information. Two different scenarios are proposed as validation with two different wearable systems. First, a custom high-performance wearable system is built and used in a free environment. A second dataset is acquired from an Android-based commercial device in a ‘wild’ scenario with rough terrains, adversarial conditions, crowded places and obstacles. Results on both systems and datasets are very promising, reducing the verification error rates by an order of magnitude with respect to the state-of-the-art technologies.

Abstract: In this paper we describe a framework to learn a model of shape variability in a set of patterns. The framework is based on the Active Appearance Model (AAM) and permits to combine shape deformations with appearance variability. We have used two modifications of the Blurred Shape Model (BSM) descriptor as basic shape and appearance features to learn the model. These modifications permit to overcome the rigidity of the original BSM, adapting it to the deformations of the shape to be represented. We have applied this framework to representation and classification of handwritten digits and symbols. We show that results of the proposed methodology outperform the original BSM approach.

Abstract: In this paper we improve the histogram of oriented gradients (HOG), a core descriptor of state-of-the-art object detection, by the use of higher-level information coming from image segmentation. The idea is to re-weight the descriptor while computing it without increasing its size. The benefits of the proposal are two-fold: (i) to improve the performance of the detector by enriching the descriptor information and (ii) take advantage of the information of image segmentation, which in fact is likely to be used in other stages of the detection system such as candidate generation or refinement.
We test our technique in the INRIA person dataset, which was originally developed to test HOG, embedding it in a human detection system. The well-known segmentation method, mean-shift (from smaller to larger super-pixels), and different methods to re-weight the original descriptor (constant, region-luminance, color or texture-dependent) has been evaluated. We achieve performance improvements of 4:47% in detection rate through the use of differences of color between contour pixel neighborhoods as re-weighting function.

Abstract: In this paper we present a multipage administrative document image retrieval system based on textual and visual representations of document pages. Individual pages are represented by textual or visual information using a bag-of-words framework. Different fusion strategies are evaluated which allow the system to perform multipage document retrieval on the basis of a single page retrieval system. Results are reported on a large dataset of document images sampled from a banking workflow.

Abstract: In this paper we propose an approach for word spotting in handwritten document images. We state the problem from a focused retrieval perspective, i.e. locating instances of a query word in a large scale dataset of digitized manuscripts. We combine two approaches, namely one based on word segmentation and another one segmentation-free. The first approach uses a hashing strategy to coarsely prune word images that are unlikely to be instances of the query word. This process is fast but has a low precision due to the errors introduced in the segmentation step. The regions containing candidate words are sent to the second process based on a state of the art technique from the visual object detection field. This discriminative model represents the appearance of the query word and computes a similarity score. In this way we propose a coarse-to-fine approach achieving a compromise between efficiency and accuracy. The validation of the model is shown using a collection of old handwritten manuscripts. We appreciate a substantial improvement in terms of precision regarding the previous proposed method with a low computational cost increase.

Abstract: In this paper, we address the problem of human action recognition in reconstructed 3-D data acquired by multi-camera systems. We contribute to this field by introducing a novel 3-D action recognition approach based on detection of 4-D (3-D space $+$ time) spatio-temporal interest points (STIPs) and local description of 3-D motion features. STIPs are detected in multi-view images and extended to 4-D using 3-D reconstructions of the actors and pixel-to-vertex correspondences of the multi-camera setup. Local 3-D motion descriptors, histogram of optical 3-D flow (HOF3D), are extracted from estimated 3-D optical flow in the neighborhood of each 4-D STIP and made view-invariant. The local HOF3D descriptors are divided using 3-D spatial pyramids to capture and improve the discrimination between arm- and leg-based actions. Based on these pyramids of HOF3D descriptors we build a bag-of-words (BoW) vocabulary of human actions, which is compressed and classified using agglomerative information bottleneck (AIB) and support vector machines (SVMs), respectively. Experiments on the publicly available i3DPost and IXMAS datasets show promising state-of-the-art results and validate the performance and view-invariance of the approach.