Abstract: Visual Question Answering (VQA) is the task where given an image and a natural language question, the objective is to generate a natural language answer. At the intersection between computer vision and natural language processing, this task can be seen as a measure of image understanding capabilities, as it requires to reason about objects, actions, colors, positions, the relations between the different elements as well as commonsense reasoning, world knowledge, arithmetic skills and natural language understanding. However, even though the text present in the images conveys important semantically rich information that is explicit and not available in any other form, most VQA methods remained illiterate, largely
ignoring the text despite its potential significance. In this thesis, we set out on a journey to bring reading capabilities to computer vision models applied to the VQA task, creating new datasets and methods that can read, reason and integrate the text with other visual cues in natural scene images and documents.
In Chapter 3, we address the combination of scene text with visual information to fully understand all the nuances of natural scene images. To achieve this objective, we define a new sub-task of VQA that requires reading the text in the image, and highlight the limitations of the current methods. In addition, we propose a new architecture that integrates both modalities and jointly reasons about textual and visual features. In Chapter 5, we shift the domain of VQA with reading capabilities and apply it on scanned industry document images, providing a high-level end-purpose perspective to Document Understanding, which has been
primarily focused on digitizing the document’s contents and extracting key values without considering the ultimate purpose of the extracted information. For this, we create a dataset which requires methods to reason about the unique and challenging elements of documents, such as text, images, tables, graphs and complex layouts, to provide accurate answers in natural language. However, we observed that explicit visual features provide a slight contribution in the overall performance, since the main information is usually conveyed within the text and its position. In consequence, in Chapter 6, we propose VQA on infographic images, seeking for document images with more visually rich elements that require to fully exploit visual information in order to answer the questions. We show the performance gap of
different methods when used over industry scanned and infographic images, and propose a new method that integrates the visual features in early stages, which allows the transformer architecture to exploit the visual features during the self-attention operation. Instead, in Chapter 7, we apply VQA on a big collection of single-page documents, where the methods must find which documents are relevant to answer the question, and provide the answer itself. Finally, in Chapter 8, mimicking real-world application problems where systems must process documents with multiple pages, we address the multipage document visual question answering task. We demonstrate the limitations of existing methods, including models specifically designed to process long sequences. To overcome these limitations, we propose
a hierarchical architecture that can process long documents, answer questions, and provide the index of the page where the information to answer the question is located as an explainability measure.

Abstract: From its early stages, the community of Pattern Recognition and Computer Vision has considered the importance of leveraging the structural information when understanding images. Usually, graphs have been proposed as a suitable model to represent this kind of information due to their flexibility and representational power able to codify both, the components, objects, or entities and their pairwise relationship. Even though graphs have been successfully applied to a huge variety of tasks, as a result of their symbolic and relational nature, graphs have always suffered from some limitations compared to statistical approaches. Indeed, some trivial mathematical operations do not have an equivalence in the graph domain. For instance, in the core of many pattern recognition applications, there is a need to compare two objects. This operation, which is trivial when considering feature vectors defined in \(\mathbb{R}^n\), is not properly defined for graphs.

In this thesis, we have investigated the importance of the structural information from two perspectives, the traditional graph-based methods and the new advances on Geometric Deep Learning. On the one hand, we explore the problem of defining a graph representation and how to deal with it on a large scale and noisy scenario. On the other hand, Graph Neural Networks are proposed to first redefine a Graph Edit Distance methodologies as a metric learning problem, and second, to apply them in a real use case scenario for the detection of repetitive patterns which define tables in invoice documents. As experimental framework, we have validated the different methodological contributions in the domain of Document Image Analysis and Recognition.

Abstract: Optical Flow (OF) is the input of a wide range of decision support systems such as car driver assistance, UAV guiding or medical diagnose. In these real situations, the absence of ground truth forces to assess OF quality using quantities computed from either sequences or the computed optical flow itself. These quantities are generally known as Confidence Measures, CM. Even if we have a proper confidence measure we still need a way to evaluate its ability to discard pixels with an OF prone to have a large error. Current approaches only provide a descriptive evaluation of the CM performance but such approaches are not capable to fairly compare different confidence measures and optical flow algorithms. Thus, it is of prime importance to define a framework and a general road map for the evaluation of optical flow performance.
This thesis provides a framework able to decide which pairs “ optical flow – confidence measure” (OF-CM) are best suited for optical flow error bounding given a confidence level determined by a decision support system. To design this framework we cover the following points:
Descriptive scores. As a first step, we summarize and analyze the sources of inaccuracies in the output of optical flow algorithms. Second, we present several descriptive plots that visually assess CM capabilities for OF error bounding. In addition to the descriptive plots, given a plot representing OF-CM capabilities to bound the error, we provide a numeric score that categorizes the plot according to its decreasing profile, that is, a score assessing CM performance.
Statistical framework. We provide a comparison framework that assesses the best suited OF-CM pair for error bounding that uses a two stage cascade process. First of all we assess the predictive value of the confidence measures by means of a descriptive plot. Then, for a sample of descriptive plots computed over training frames, we obtain a generic curve that will be used for sequences with no ground truth. As a second step, we evaluate the obtained general curve and its capabilities to really reflect the predictive value of a confidence measure using the variability across train frames by means of ANOVA.
The presented framework has shown its potential in the application on clinical decision support systems. In particular, we have analyzed the impact of the different image artifacts such as noise and decay to the output of optical flow in a cardiac diagnose system and we have improved the navigation inside the bronchial tree on bronchoscopy.

Abstract: Single image super-resolution (SISR) is an important task in image processing which aims to enhance the resolution of imaging systems. Recently, SISR has witnessed great strides with the rapid development of deep learning. Recent advances in SISR are mostly devoted to designing deeper and wider networks to enhance their representation learning capacity. However, as the depth of networks increases, deep learning-based methods are faced with the challenge of computational complexity in practice. Moreover, most existing methods rarely leverage the intermediate features and also do not discriminate the computation of features by their frequencial components, thereby achieving relatively low performance. Aside from the aforementioned problems, another desired ability is to upsample images to arbitrary scales using a single model. Most current SISR methods train a dedicated model for each target resolution, losing generality and increasing memory requirements. In this thesis, we address the aforementioned issues and propose solutions to them: i) We present a novel frequency-based enhancement block which treats different frequencies in a heterogeneous way and also models inter-channel dependencies, which consequently enrich the output feature. Thus it helps the network generate more discriminative representations by explicitly recovering finer details. ii) We introduce OverNet which contains two main parts: a lightweight feature extractor that follows a novel recursive framework of skip and dense connections to reduce low-level feature degradation, and an overscaling module that generates an accurate SR image by internally constructing an overscaled intermediate representation of the output features. Then, to solve the problem of reconstruction at arbitrary scale factors, we introduce a novel multi-scale loss, that allows the simultaneous training of all scale factors using a single model. iii) We propose a directional variance attention network which leverages a novel attention mechanism to enhance features in different channels and spatial regions. Moreover, we introduce a novel procedure for using attention mechanisms together with residual blocks to facilitate the preservation of finer details. Finally, we demonstrate that our approaches achieve considerably better performance than previous state-of-the-art methods, in terms of both quantitative and visual quality.

Abstract: Eye tracker devices have traditionally been only used inside laboratories, requiring trained professionals and elaborate setup mechanisms. However, in the recent years the scientific work on easier–to–use eye trackers which require no special hardware—other than the omnipresent front facing cameras in computers, tablets, and mobiles—is aiming at making this technology common–place. These types of trackers have several extra challenges that make the problem harder, such as low resolution images provided by a regular webcam, the changing ambient lighting conditions, personal appearance differences, changes in head pose, and so on. Recent research in the field has focused on all these challenges in order to provide better gaze estimation performances in a real world setup.
In this work, we aim at tackling the gaze tracking problem in a single camera setup. We first analyze all the previous work in the field, identifying the strengths and weaknesses of each tried idea. We start our work on the gaze tracker with an appearance–based gaze estimation method, which is the simplest idea that creates a direct mapping between a rectangular image patch extracted around the eye in a camera image, and the gaze point (or gaze direction). Here, we do an extensive analysis of the factors that affect the performance of this tracker in several experimental setups, in order to address these problems in future works. In the second part of our work, we propose a feature–based gaze estimation method, which encodes the eye region image into a compact representation. We argue that this type of representation is better suited to dealing with head pose and lighting condition changes, as it both reduces the dimensionality of the input (i.e. eye image) and breaks the direct connection between image pixel intensities and the gaze estimation. Lastly, we use a face alignment algorithm to have robust face pose estimation, using a 3D model customized to the subject using the tracker. We combine this with a convolutional neural network trained on a large dataset of images to build a face pose invariant gaze tracker.

Abstract: Motion perception is one of the most important attributes of the human brain. Visual motion perception consists in inferring speed and direction of elements in a scene based on visual inputs. Analogously, computer vision is assisted by motion cues in the scene. Motion detection in computer vision is useful in solving problems such as segmentation, depth from motion, structure from motion, compression, navigation and many others. These problems are common in several applications, for instance, video surveillance, robot navigation and advanced driver assistance systems (ADAS). One of the most widely used techniques for motion detection is the optical flow estimation. The work in this thesis attempts to make optical flow suitable for the requirements and conditions of driving scenarios. In this context, a novel space-variant representation called reverse log-polar representation is proposed that is shown to be better than the traditional log-polar space-variant representation for ADAS. The space-variant representations reduce the amount of data to be processed. Another major contribution in this research is related to the analysis of the influence of specific characteristics from driving scenarios on the optical flow accuracy. Characteristics such as vehicle speed and
road texture are considered in the aforementioned analysis. From this study, it is inferred that the regularization weight has to be adapted according to the required error measure and for different speeds and road textures. It is also shown that polar represented optical flow suits driving scenarios where predominant motion is translation. Due to the requirements of such a study and by the lack of needed datasets a new synthetic dataset is presented; it contains: i) sequences of different speeds and road textures in an urban scenario; ii) sequences with complex motion of an on-board camera; and iii) sequences with additional moving vehicles in the scene. The ground-truth optical flow is generated by the ray-tracing technique. Further, few applications of optical flow in ADAS are shown. Firstly, a robust RANSAC based technique to estimate horizon line is proposed. Then, an egomotion estimation is presented to compare the proposed space-variant representation with the classical one. As a final contribution, a modification in the regularization term is proposed that notably improves the results
in the ADAS applications. This adaptation is evaluated using a state of the art optical flow technique. The experiments on a public dataset (KITTI) validate the advantages of using the proposed modification.