Abstract: Most popular metric learning losses have no direct relation with the evaluation metrics that are subsequently applied to evaluate their performance. We hypothesize that training a metric learning model by maximizing the area under the ROC curve (which is a typical performance measure of recognition systems) can induce an implicit ranking suitable for retrieval problems. This hypothesis is supported by previous work that proved that a curve dominates in ROC space if and only if it dominates in Precision-Recall space. To test this hypothesis, we design and maximize an approximated, derivable relaxation of the area under the ROC curve. The proposed AUC loss achieves state-of-the-art results on two large scale retrieval benchmark datasets (Stanford Online Products and DeepFashion In-Shop). Moreover, the AUC loss achieves comparable performance to more complex, domain specific, state-of-the-art methods for vehicle re-identification.

Abstract: Recent self-supervised learning methods are able to learn high-quality image representations and are closing the gap with supervised approaches. However, these methods are unable to acquire new knowledge incrementally – they are, in fact, mostly used only as a pre-training phase over IID data. In this work we investigate self-supervised methods in continual learning regimes without any replay
mechanism. We show that naive functional regularization,also known as feature distillation, leads to lower plasticity and limits continual learning performance. Instead, we propose Projected Functional Regularization in which a separate temporal projection network ensures that the newly learned feature space preserves information of the previous one, while at the same time allowing for the learning of new features. This prevents forgetting while maintaining the plasticity of the learner. Comparison with other incremental learning approaches applied to self-supervision demonstrates that our method obtains competitive performance in
different scenarios and on multiple datasets.

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: The mass irruption of Deep Convolutional Neural Networks (CNNs) in computer vision since 2012 led to a dominance of the image understanding paradigm consisting in an end-to-end fully supervised learning workflow over large-scale annotated datasets. This approach proved to be extremely useful at solving a myriad of classic and new computer vision tasks with unprecedented performance —often, surpassing that of humans—, at the expense of vast amounts of human-labeled data, extensive computational resources and the disposal of all of our prior knowledge on the task at hand. Even though simple transfer learning methods, such as fine-tuning, have achieved remarkable impact, their success when the amount of labeled data in the target domain is small is limited. Furthermore, the non-static nature of data generation sources will often derive in data distribution shifts that degrade the performance of deployed models. As a consequence, there is a growing demand for methods that can exploit elements of prior knowledge and sources of information other than the manually generated ground truth annotations of the images during the network training process, so that they can adapt to new domains that constitute, if not a small data regime, at least a small labeled data regime. This thesis targets such few or no labeled data scenario in three distinct image-to-image mapping learning problems. It contributes with various approaches that leverage our previous knowledge of different elements of the image formation process: We first present a data-efficient framework for both defocus and motion blur detection, based on a model able to produce realistic synthetic local degradations. The framework comprises a self-supervised, a weakly-supervised and a semi-supervised instantiation, depending on the absence or availability and the nature of human annotations, and outperforms fully-supervised counterparts in a variety of settings. Our knowledge on color image formation is then used to gather input and target ground truth image pairs for the RGB to hyperspectral image reconstruction task. We make use of a CNN to tackle this problem, which, for the first time, allows us to exploit spatial context and achieve state-of-the-art results given a limited hyperspectral image set. In our last contribution to the subfield of data-efficient image-to-image transformation problems, we present the novel semi-supervised task of zero-pair cross-view semantic segmentation: we consider the case of relocation of the camera in an end-to-end trained and deployed monocular, fixed-view semantic segmentation system often found in industry. Under the assumption that we are allowed to obtain an additional set of synchronized but unlabeled image pairs of new scenes from both original and new camera poses, we present ZPCVNet, a model and training procedure that enables the production of dense semantic predictions in either source or target views at inference time. The lack of existing suitable public datasets to develop this approach led us to the creation of MVMO, a large-scale Multi-View, Multi-Object path-traced dataset with per-view semantic segmentation annotations. We expect MVMO to propel future research in the exciting under-developed fields of cross-view and multi-view semantic segmentation. Last, in a piece of applied research of direct application in the context of process monitoring of an Electric Arc Furnace (EAF) in a steelmaking plant, we also consider the problem of simultaneously estimating the temperature and spectral emissivity of distant hot emissive samples. To that end, we design our own capturing device, which integrates three point spectrometers covering a wide range of the Ultra-Violet, visible, and Infra-Red spectra and is capable of registering the radiance signal incoming from an 8cm diameter spot located up to 20m away. We then define a physically accurate radiative transfer model that comprises the effects of atmospheric absorbance, of the optical system transfer function, and of the sample temperature and spectral emissivity themselves. We solve this inverse problem without the need for annotated data using a probabilistic programming-based Bayesian approach, which yields full posterior distribution estimates of the involved variables that are consistent with laboratory-grade measurements.

Abstract: In most eukaryotes, pairing of homologous chromosomes is an essential feature of meiosis that ensures homologous recombination and segregation. However, when the pairing process begins, it is still under investigation. Contrasting data exists in Mus musculus, since both leptotene DSB-dependent and preleptotene DSB-independent mechanisms have been described. To unravel this contention, we examined homologous pairing in pre-meiotic and meiotic Mus musculus cells using a threedimensional fuorescence in situ hybridization-based protocol, which enables the analysis of the entire karyotype using DNA painting probes. Our data establishes in an unambiguously manner that 73.83% of homologous chromosomes are already paired at premeiotic stages (spermatogonia-early preleptotene spermatocytes). The percentage of paired homologous chromosomes increases to 84.60% at mid-preleptotene-zygotene stage, reaching 100% at pachytene stage. Importantly, our results demonstrate a high percentage of homologous pairing observed before the onset of meiosis; this pairing does not occur randomly, as the percentage was higher than that observed in somatic cells (19.47%) and between nonhomologous chromosomes (41.1%). Finally, we have also observed that premeiotic homologous pairing is asynchronous and independent of the chromosome size, GC content, or presence of NOR regions.

Abstract: Handwritten Music Recognition, especially in the historical domain, is an inherently challenging endeavour; paper degradation artefacts and the ambiguous nature of handwriting make recognising such scores an error-prone process, even for the current state-of-the-art Sequence to Sequence models. In this work we propose a way of reducing the production of statistically implausible output sequences by fusing a Language Model into a recognition Sequence to Sequence model. The idea is leveraging visually-conditioned and context-conditioned output distributions in order to automatically find and correct any mistakes that would otherwise break context significantly. We have found this approach to improve recognition results to 25.15 SER (%) from a previous best of 31.79 SER (%) in the literature.

Abstract: In class-incremental semantic segmentation, we have no access to the labeled data of previous tasks. Therefore, when incrementally learning new classes, deep neural networks suffer from catastrophic forgetting of previously learned knowledge. To address this problem, we propose to apply a self-training approach that leverages unlabeled data, which is used for rehearsal of previous knowledge. Specifically, we first learn a temporary model for the current task, and then, pseudo labels for the unlabeled data are computed by fusing information from the old model of the previous task and the current temporary model. In addition, conflict reduction is proposed to resolve the conflicts of pseudo labels generated from both the old and temporary models. We show that maximizing self-entropy can further improve results by smoothing the overconfident predictions. Interestingly, in the experiments, we show that the auxiliary data can be different from the training data and that even general-purpose, but diverse auxiliary data can lead to large performance gains. The experiments demonstrate the state-of-the-art results: obtaining a relative gain of up to 114% on Pascal-VOC 2012 and 8.5% on the more challenging ADE20K compared to previous state-of-the-art methods.

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.