Abstract: Vision and Language are broadly regarded as cornerstones of intelligence. Even though language and vision have different aims – language having the purpose of communication, transmission of information and vision having the purpose of constructing mental representations around us to navigate and interact with objects – they cooperate and depend on one another in many tasks we perform effortlessly. This reliance is actively being studied in various Computer Vision tasks, e.g. image captioning, visual question answering, image-sentence retrieval, phrase grounding, just to name a few. All of these tasks share the inherent difficulty of the aligning the two modalities, while being robust to language
priors and various biases existing in the datasets. One of the ultimate goal for vision and language research is to be able to inject world knowledge while getting rid of the biases that come with the datasets. In this thesis, we mainly focus on two vision and language tasks, namely Image Captioning and Scene-Text Visual Question Answering (STVQA).
In both domains, we start by defining a new task that requires the utilization of world knowledge and in both tasks, we find that the models commonly employed are prone to biases that exist in the data. Concretely, we introduce new tasks and discover several problems that impede performance at each level and provide remedies or possible solutions in each chapter: i) We define a new task to move beyond Image Captioning to Image Interpretation that can utilize Named Entities in the form of world knowledge. ii) We study the object hallucination problem in classic Image Captioning systems and develop an architecture-agnostic solution. iii) We define a sub-task of Visual Question Answering that requires reading the text in the image (STVQA), where we highlight the limitations of current models. iv) We propose an architecture for the STVQA task that can point to the answer in the image and show how to combine it with classic VQA models. v) We show how far language can get us in STVQA and discover yet another bias which causes the models to disregard the image while doing Visual Question Answering.

Abstract: In this thesis, we propose different contributions with the goal of enhancing and recognizing historical handwritten document images, especially the ones with rare scripts, such as cipher documents.
In the first part, some effective end-to-end models for Document Image Enhancement (DIE) using deep learning models were presented. First, Generative Adversarial Networks (cGAN) for different tasks (document clean-up, binarization, deblurring, and watermark removal) were explored. Next, we further improve the results by recovering the degraded document images into a clean and readable form by integrating a text recognizer into the cGAN model to promote the generated document image to be more readable. Afterward, we present a new encoder-decoder architecture based on vision transformers to enhance both machine-printed and handwritten document images, in an end-to-end fashion.
The second part of the thesis addresses Handwritten Text Recognition (HTR) in low resource scenarios, i.e. when only few labeled training data is available. We propose novel methods for recognizing ciphers with rare scripts. First, a few-shot object detection based method was proposed. Then, we incorporate a progressive learning strategy that automatically assignspseudo-labels to a set of unlabeled data to reduce the human labor of annotating few pages while maintaining the good performance of the model. Secondly, a data generation technique based on Bayesian Program Learning (BPL) is proposed to overcome the lack of data in such rare scripts. Thirdly, we propose a Text-Degradation Invariant Auto Encoder (Text-DIAE). This latter self-supervised model is designed to tackle two tasks, text recognition and document image enhancement. The proposed model does not exhibit limitations of previous state-of-the-art methods based on contrastive losses, while at the same time, it requires substantially fewer data samples to converge.
In the third part of the thesis, we analyze, from the user perspective, the usage of HTR systems in low resource scenarios. This contrasts with the usual research on HTR, which often focuses on technical aspects only and rarely devotes efforts on implementing software tools for scholars in Humanities.

Abstract: Deep learning has made tremendous progress in the last decade due to the explosion of training data and computational power. Through end-to-end training on a
large dataset, image representations are more discriminative than the previously
used hand-crafted features. However, for many real-world applications, training
and testing on a single dataset is not realistic, as the test distribution may change over time. Continuous learning takes this situation into account, where the learner must adapt to a sequence of tasks, each with a different distribution. If you would naively continue training the model with a new task, the performance of the model would drop dramatically for the previously learned data. This phenomenon is known as catastrophic forgetting.
Many approaches have been proposed to address this problem, which can be divided into three main categories: regularization-based approaches, rehearsal-based
approaches, and parameter isolation-based approaches. However, most of the existing works focus on image classification tasks and many other computer vision tasks
have not been well-explored in the continual learning setting. Therefore, in this
thesis, we study continual learning for image generation, object re-identification,
and object counting.
For the image generation problem, since the model can generate images from the previously learned task, it is free to apply rehearsal without any limitation. We developed two methods based on generative replay. The first one uses the generated image for joint training together with the new data. The second one is based on
output pixel-wise alignment. We extensively evaluate these methods on several
benchmarks.
Next, we study continual learning for object Re-Identification (ReID). Although
most state-of-the-art methods of ReID and continual ReID use softmax-triplet loss,
we found that it is better to solve the ReID problem from a meta-learning perspective because continual learning of reID can benefit a lot from the generalization of metalearning. We also propose a distillation loss and found that the removal of the positive pairs before the distillation loss is critical.
Finally, we study continual learning for the counting problem. We study the mainstream method based on density maps and propose a new approach for density
map distillation. We found that fixing the counter head is crucial for the continual learning of object counting. To further improve results, we propose an adaptor to adapt the changing feature extractor for the fixed counter head. Extensive evaluation shows that this results in improved continual learning performance.

Abstract: Though they achieve great success, deep neural networks typically require a huge
amount of labeled data for training. However, collecting labeled data is often laborious and expensive. It would, therefore, be ideal if the knowledge obtained from label-rich datasets could be transferred to unlabeled data. However, deep networks are weak at generalizing to unseen domains, even when the differences are only subtle between the datasets. In real-world situations, a typical factor impairing the model generalization ability is the distribution shift between data from different domains, which is a long-standing problem usually termed as (unsupervised) domain adaptation.
A crucial requirement in the methodology of these domain adaptation methods is that they require access to source domain data during the adaptation process to the target domain. Accessibility to the source data of a trained source model is often impossible in real-world applications, for example, when deploying domain adaptation algorithms on mobile devices where the computational capacity is limited or in situations where data privacy rules limit access to the source domain data. Without access to the source domain data, existing methods suffer from inferior performance. Thus, in this thesis, we investigate domain adaptation without source data (termed as source-free domain adaptation) in multiple different scenarios that focus on image classification tasks.
We first study the source-free domain adaptation problem in a closed-set setting,
where the label space of different domains is identical. Only accessing the pretrained source model, we propose to address source-free domain adaptation from the perspective of unsupervised clustering. We achieve this based on nearest neighborhood clustering. In this way, we can transfer the challenging source-free domain adaptation task to a type of clustering problem. The final optimization objective is an upper bound containing only two simple terms, which can be explained as discriminability and diversity. We show that this allows us to relate several other methods in domain adaptation, unsupervised clustering and contrastive learning via the perspective of discriminability and diversity.

Abstract: This thesis embarks on an exploratory journey into robustness in deep learning,
with a keen focus on the intertwining facets of generalization, explainability, and
edge cases within the realm of computer vision. In deep learning, robustness
epitomizes a model’s resilience and flexibility, grounded on its capacity to generalize across diverse data distributions, explain its predictions transparently, and navigate the intricacies of edge cases effectively. The challenges associated with robust generalization are multifaceted, encompassing the model’s performance on unseen data and its defense against out-of-distribution data and adversarial attacks. Bridging this gap, the potential of Embedding Propagation (EP) for improving out-of-distribution generalization is explored. EP is depicted as a powerful tool facilitating manifold smoothing, which in turn fortifies the model’s robustness against adversarial onslaughts and bolsters performance in few-shot and self-/semi-supervised learning scenarios. In the labyrinth of deep learning models, the path to robustness often intersects with explainability. As model complexity increases, so does the urgency to decipher their decision-making
processes. Acknowledging this, the thesis introduces a robust framework for
evaluating and comparing various counterfactual explanation methods, echoing
the imperative of explanation quality over quantity and spotlighting the intricacies of diversifying explanations. Simultaneously, the deep learning landscape is fraught with edge cases – anomalies in the form of small objects or rare instances in object detection tasks that defy the norm. Confronting this, the
thesis presents an extension of the DETR (DEtection TRansformer) model to enhance small object detection. The devised DETR-FP, embedding the Feature Pyramid technique, demonstrating improvement in small objects detection accuracy, albeit facing challenges like high computational costs. With emergence of foundation models in mind, the thesis unveils EarthView, the largest scale remote sensing dataset to date, built for the self-supervised learning of a robust foundational model for remote sensing. Collectively, these studies contribute to the grand narrative of robustness in deep learning, weaving together the strands of generalization, explainability, and edge case performance. Through these methodological advancements and novel datasets, the thesis calls for continued exploration, innovation, and refinement to fortify the bastion of robust computer vision.

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: Action recognition is a very challenging and important problem in computer vi­sion. Researchers working on this field aspire to provide computers with the abil­ ity to visually perceive human actions – that is, to observe, interpret, and under­ stand human-related events that occur in the physical environment merely from visual data. The applications of this technology are numerous: human-machine interaction, e-health, monitoring/surveillance, and content-based video retrieval, among others. Hand-crafted methods dominated the field until the apparition of the first successful deep learning-based action recognition works. Although ear­ lier deep-based methods underperformed with respect to hand-crafted approaches, these slowly but steadily improved to become state-of-the-art, eventually achieving better results than hand-crafted ones. Still, hand-crafted approaches can be advan­ tageous in certain scenarios, specially when not enough data is available to train very large deep models or simply to be combined with deep-based methods to fur­ ther boost the performance. Hence, showing how hand-crafted features can provide extra knowledge the deep networks are notable to easily learn about human actions.
This Thesis concurs in time with this change of paradigm and, hence, reflects it into two distinguished parts. In the first part, we focus on improving current suc­ cessful hand-crafted approaches for action recognition and we do so from three dif­ ferent perspectives. Using the dense trajectories framework as a backbone: first, we explore the use of multi-modal and multi-view input
data to enrich the trajectory de­ scriptors. Second, we focus on the classification part of action recognition pipelines and propose an ensemble learning approach, where each classifier leams from a dif­ferent set of local spatiotemporal features to then combine their outputs following an strategy based on the Dempster-Shaffer Theory. And third, we propose a novel hand-crafted feature extraction method that constructs a rnid-level feature descrip­ tion to better modellong-term spatiotemporal dynarnics within action videos. Moving to the second part of the Thesis, we start with a comprehensive study of the current deep-learning based action recognition methods. We review both fun­ damental and cutting edge methodologies reported during the last few years and introduce a taxonomy of deep-leaming methods dedicated to action recognition. In particular, we analyze and discuss how these handle
the temporal dimension of data. Last but not least, we propose a residual recurrent network for action recogni­ tion that naturally integrates all our previous findings in a powerful and prornising framework.