Abstract: The curse of data labeling is a costly bottleneck in supervised deep learning, where large amounts of labeled data are needed to train intelligent systems. In onboard perception for autonomous driving, this cost corresponds to the labeling of raw data from sensors such as cameras, LiDARs, RADARs, etc. Therefore, synthetic data with automatically generated ground truth (labels) has aroused as a reliable alternative for training onboard perception models.
However, synthetic data commonly suffers from synth-to-real domain shift, i.e., models trained on the synthetic domain do not show their achievable accuracy when performing in the real world. This shift needs to be addressed by techniques falling in the realm of domain adaptation (DA).
The semi-supervised learning (SSL) paradigm can be followed to address DA. In this case, a model is trained using source data with labels (here synthetic) and leverages minimal knowledge from target data (here the real world) to generate pseudo-labels. These pseudo-labels help the training process to reduce the gap between the source and the target domains. In general, we can assume accessing both, pseudo-labels and a few amounts of human-provided labels for the target-domain data. However, the most interesting and challenging setting consists in assuming that we do not have human-provided labels at all. This setting is known as unsupervised domain adaptation (UDA). This PhD focuses on applying SSL to the UDA setting, for onboard visual tasks related to autonomous driving. We start by addressing the synth-to-real UDA problem on onboard vision-based object detection (pedestrians and cars), a critical task for autonomous driving and driving assistance. In particular, we propose to apply an SSL technique known as co-training, which we adapt to work with deep models that process a multi-modal input. The multi-modality consists of the visual appearance of the images (RGB) and their monocular depth estimation. The synthetic data we use as the source domain contains both, object bounding boxes and depth information. This prior knowledge is the
starting point for the co-training technique, which iteratively labels unlabeled real-world data and uses such pseudolabels (here bounding boxes with an assigned object class) to progressively improve the labeling results. Along this
process, two models collaborate to automatically label the images, in a way that one model compensates for the errors of the other, so avoiding error drift. While this automatic labeling process is done offline, the resulting pseudolabels can be used to train object detection models that must perform in real-time onboard a vehicle. We show that multi-modal co-training improves the labeling results compared to single-modal co-training, remaining competitive compared to human labeling.
Given the success of co-training in the context of object detection, we have also adapted this technique to a more crucial and challenging visual task, namely, onboard semantic segmentation. In fact, providing labels for a single image
can take from 30 to 90 minutes for a human labeler, depending on the content of the image. Thus, developing automatic labeling techniques for this visual task is of great interest to the automotive industry. In particular, the new co-training framework addresses synth-to-real UDA by an initial stage of self-training. Intermediate models arising from this stage are used to start the co-training procedure, for which we have elaborated an accurate collaboration policy between the two models performing the automatic labeling. Moreover, our co-training seamlessly leverages datasets from different synthetic domains. In addition, the co-training procedure is agnostic to the loss function used to train the semantic segmentation models which perform the automatic labeling. We achieve state-of-the-art results on publicly available benchmark datasets, again, remaining competitive compared to human labeling.
Finally, on the ground of our previous experience, we have designed and implemented a new SSL technique for UDA in the context of visual semantic segmentation. In this case, we mimic the labeling methodology followed by human labelers. In particular, rather than labeling full images at a time, categories of semantic classes are defined and only those are labeled in a labeling pass. In fact, different human labelers can become specialists in labeling different categories. Afterward, these per-category-labeled layers are combined to provide fully labeled images. Our technique is inspired by this methodology since we perform synth-to-real UDA per category, using the self-training stage previously developed as part of our co-training framework. The pseudo-labels obtained for each category are finally
fused to obtain fully automatically labeled images. In this context, we have also contributed to the development of a new photo-realistic synthetic dataset based on path-tracing rendering. Our new SSL technique seamlessly leverages publicly available synthetic datasets as well as this new one to obtain state-of-the-art results on synth-to-real UDA for semantic segmentation. We show that the new dataset allows us to reach better labeling accuracy than previously existing datasets, at the same time that it complements well them when combined. Moreover, we also show that the new human-inspired SSL technique outperforms co-training.

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: Self-supervised learning (SSL) is a powerful technique for learning robust representations from unlabeled data. By learning to remain invariant to applied data augmentations, methods such as SimCLR and MoCo are able to reach quality on par with supervised approaches. However, this invariance may be harmful to solving some downstream tasks which depend on traits affected by augmentations used during pretraining, such as color. In this paper, we propose to foster sensitivity to such characteristics in the representation space by modifying the projector network, a common component of self-supervised architectures. Specifically, we supplement the projector with information about augmentations applied to images. In order for the projector to take advantage of this auxiliary conditioning when solving the SSL task, the feature extractor learns to preserve the augmentation information in its representations. Our approach, coined Conditional Augmentation-aware Self-supervised Learning (CASSLE), is directly applicable to typical joint-embedding SSL methods regardless of their objective functions. Moreover, it does not require major changes in the network architecture or prior knowledge of downstream tasks. In addition to an analysis of sensitivity towards different data augmentations, we conduct a series of experiments, which show that CASSLE improves over various SSL methods, reaching state-of-the-art performance in multiple downstream tasks.

Abstract: CodaLab Competitions is an open source web platform designed to help data scientists and research teams to crowd-source the resolution of machine learning problems through the organization of competitions, also called challenges or contests. CodaLab Competitions provides useful features such as multiple phases, results and code submissions, multi-score leaderboards, and jobs running
inside Docker containers. The platform is very flexible and can handle large scale experiments, by allowing organizers to upload large datasets and provide their own CPU or GPU compute workers.

Abstract: This paper presents two algorithms for the large-scale automatic detection and instance segmentation of potential archaeological mounds on historical maps. Historical maps present a unique source of information for the reconstruction of ancient landscapes. The last 100 years have seen unprecedented landscape modifications with the introduction and large-scale implementation of mechanised agriculture, channel-based irrigation schemes, and urban expansion to name but a few. Historical maps offer a window onto disappearing landscapes where many historical and archaeological elements that no longer exist today are depicted. The algorithms focus on the detection and shape extraction of mound features with high probability of being archaeological settlements, mounds being one of the most commonly documented archaeological features to be found in the Survey of India historical map series, although not necessarily recognised as such at the time of surveying. Mound features with high archaeological potential are most commonly depicted through hachures or contour-equivalent form-lines, therefore, an algorithm has been designed to detect each of those features. Our proposed approach addresses two of the most common issues in archaeological automated survey, the low-density of archaeological features to be detected, and the small amount of training data available. It has been applied to all types of maps available of the historic 1″ to 1-mile series, thus increasing the complexity of the detection. Moreover, the inclusion of synthetic data, along with a Curriculum Learning strategy, has allowed the algorithm to better understand what the mound features look like. Likewise, a series of filters based on topographic setting, form, and size have been applied to improve the accuracy of the models. The resulting algorithms have a recall value of 52.61% and a precision of 82.31% for the hachure mounds, and a recall value of 70.80% and a precision of 70.29% for the form-line mounds, which allowed the detection of nearly 6000 mound features over an area of 470,500 km2, the largest such approach to have ever been applied. If we restrict our focus to the maps most similar to those used in the algorithm training, we reach recall values greater than 60% and precision values greater than 90%. This approach has shown the potential to implement an adaptive algorithm that allows, after a small amount of retraining with data detected from a new map, a better general mound feature detection in the same map.

Abstract: Despite achieving state-of-the-art performance in visual tracking, recent single-branch trackers tend to overlook the weak prior assumptions associated with the Vision Transformer (ViT) encoder and inference pipeline. Moreover, the effectiveness of discriminative trackers remains constrained due to the adoption of the dual-branch pipeline. To tackle the inferior effectiveness of the vanilla ViT, we propose an Adaptive ViT Model Prediction tracker (AViTMP) to bridge the gap between single-branch network and discriminative models. Specifically, in the proposed encoder AViT-Enc, we introduce an adaptor module and joint target state embedding to enrich the dense embedding paradigm based on ViT. Then, we combine AViT-Enc with a dense-fusion decoder and a discriminative target model to predict accurate location. Further, to mitigate the limitations of conventional inference practice, we present a novel inference pipeline called CycleTrack, which bolsters the tracking robustness in the presence of distractors via bidirectional cycle tracking verification. Lastly, we propose a dual-frame update inference strategy that adeptively handles significant challenges in long-term scenarios. In the experiments, we evaluate AViTMP on ten tracking benchmarks for a comprehensive assessment, including LaSOT, LaSOTExtSub, AVisT, etc. The experimental results unequivocally establish that AViTMP attains state-of-the-art performance, especially on long-time tracking and robustness.