Abstract: We present a new dataset for Visual Question Answering (VQA) on document images called DocVQA. The dataset consists of 50,000 questions defined on 12,000+ document images. Detailed analysis of the dataset in comparison with similar datasets for VQA and reading comprehension is presented. We report several baseline results by adopting existing VQA and reading comprehension models. Although the existing models perform reasonably well on certain types of questions, there is large performance gap compared to human performance (94.36% accuracy). The models need to improve specifically on questions where understanding structure of the document is crucial. The dataset, code and leaderboard are available at docvqa. org

Abstract: arXiv:2007.12562
Most of the saliency methods are evaluated on their ability to generate saliency maps, and not on their functionality in a complete vision pipeline, like for instance, image classification. In the current paper, we propose an approach which does not require explicit saliency maps to improve image classification, but they are learned implicitely, during the training of an end-to-end image classification task. We show that our approach obtains similar results as the case when the saliency maps are provided explicitely. Combining RGB data with saliency maps represents a significant advantage for object recognition, especially for the case when training data is limited. We validate our method on several datasets for fine-grained classification tasks (Flowers, Birds and Cars). In addition, we show that our saliency estimation method, which is trained without any saliency groundtruth data, obtains competitive results on real image saliency benchmark (Toronto), and outperforms deep saliency models with synthetic images (SID4VAM).

Abstract: The performances of Sign Language Recognition (SLR) systems have improved considerably in recent years. However, several open challenges still need to be solved to allow SLR to be useful in practice. The research in the field is in its infancy in regards to the robustness of the models to a large diversity of signs and signers, and to fairness of the models to performers from different demographics. This work summarises the ChaLearn LAP Large Scale Signer Independent Isolated SLR Challenge, organised at CVPR 2021 with the goal of overcoming some of the aforementioned challenges. We analyse and discuss the challenge design, top winning solutions and suggestions for future research. The challenge attracted 132 participants in the RGB track and 59 in the RGB+Depth track, receiving more than 1.5K submissions in total. Participants were evaluated using a new large-scale multi-modal Turkish Sign Language (AUTSL) dataset, consisting of 226 sign labels and 36,302 isolated sign video samples performed by 43 different signers. Winning teams achieved more than 96% recognition rate, and their approaches benefited from pose/hand/face estimation, transfer learning, external data, fusion/ensemble of modalities and different strategies to model spatio-temporal information. However, methods still fail to distinguish among very similar signs, in particular those sharing similar hand trajectories.

Abstract: In this thesis, we explore how to estimate the effects of the light interacting with the scene objects from a single image. To achieve this goal, we focus on recovering intrinsic components like reflectance, shading, or light properties such as color and position using deep architectures. The success of these approaches relies on training on large and diversified image datasets. Therefore, we present several contributions on this such as: (a) a data-augmentation technique; (b) a ground-truth for an existing multi-illuminant dataset; (c) a family of synthetic datasets, SID for Surreal Intrinsic Datasets, with diversified backgrounds and coherent light conditions; and (d) a practical pipeline to create hybrid ground-truths to overcome the complexity of acquiring realistic light conditions in a massive way. In parallel with the creation of datasets, we trained different flexible encoder-decoder deep architectures incorporating physical constraints from the image formation models.
In the last part of the thesis, we apply all the previous experience to two different problems. Firstly, we create a large hybrid Doc3DShade dataset with real shading and synthetic reflectance under complex illumination conditions, that is used to train a two-stage architecture that improves the character recognition task in complex lighting conditions of unwrapped documents. Secondly, we tackle the problem of single image scene relighting by extending both, the SID dataset to present stronger shading and shadows effects, and the deep architectures to use intrinsic components to estimate new relit images.

Abstract: Deep convolutional neural networks (CNNs) have achieved superior performance in many visual recognition applications, such as image classification, detection and segmentation. Training deep CNNs requires huge amounts of labeled data, which is expensive and labor intensive to collect. Active learning is a paradigm aimed at reducing the annotation effort by training the model on actively selected
informative and/or representative samples. In this thesis we study several aspects of active learning including video object detection for autonomous driving systems, image classification on balanced and imbalanced datasets and the incorporation of self-supervised learning in active learning. We briefly describe our approach in each of these areas to reduce the labeling effort.
In chapter two we introduce a novel active learning approach for object detection in videos by exploiting temporal coherence. Our criterion is based on the estimated number of errors in terms of false positives and false negatives. Additionally, we introduce a synthetic video dataset, called SYNTHIA-AL, specially designed to evaluate active
learning for video object detection in road scenes. Finally, we show that our
approach outperforms active learning baselines tested on two outdoor datasets.
In the next chapter we address the well-known problem of over confidence in the neural networks. As an alternative to network confidence, we propose a new informativeness-based active learning method that captures the learning dynamics of neural network with a metric called label-dispersion. This metric is low when the network consistently assigns the same label to the sample during the course of training and high when the assigned label changes frequently. We show that label-dispersion is a promising predictor of the uncertainty of the network, and show on two benchmark datasets that an active learning algorithm based on label-dispersion obtains excellent results.
In chapter four, we tackle the problem of sampling bias in active learning methods on imbalanced datasets. Active learning is generally studied on balanced datasets where an equal amount of images per class is available. However, real-world datasets suffer from severe imbalanced classes, the so called longtail distribution. We argue that this further complicates the active learning process, since the imbalanced data pool can result in suboptimal classifiers. To address this problem in the context of active learning, we propose a general optimization framework that explicitly takes class-balancing into account. Results on three datasets show that the method is general (it can be combined with most existing active learning algorithms) and can be effectively applied to boost the performance of both informative and representative-based active learning methods. In addition, we show that also on balanced datasets our method generally results in a performance gain.
Another paradigm to reduce the annotation effort is self-training that learns from a large amount of unlabeled data in an unsupervised way and fine-tunes on few labeled samples. Recent advancements in self-training have achieved very impressive results rivaling supervised learning on some datasets. In the last chapter we focus on whether active learning and self supervised learning can benefit from each other.
We study object recognition datasets with several labeling budgets for the evaluations. Our experiments reveal that self-training is remarkably more efficient than active learning at reducing the labeling effort, that for a low labeling budget, active learning offers no benefit to self-training, and finally that the combination of active learning and self-training is fruitful when the labeling budget is high.

Abstract: Despite the latest advances in Deep Learning, the recognition of handwritten music scores is still a challenging endeavour. Even though the recent Sequence to Sequence(Seq2Seq) architectures have demonstrated its capacity to reliably recognise handwritten text, their performance is still far from satisfactory when applied to historical handwritten scores. Indeed, the ambiguous nature of handwriting, the non-standard musical notation employed by composers of the time and the decaying state of old paper make these scores remarkably difficult to read, sometimes even by trained humans. Thus, in this work we explore the incorporation of language models into a Seq2Seq-based architecture to try to improve transcriptions where the aforementioned unclear writing produces statistically unsound mistakes, which as far as we know, has never been attempted for this field of research on this architecture. After studying various Language Model integration techniques, the experimental evaluation on historical handwritten music scores shows a significant improvement over the state of the art, showing that this is a promising research direction for dealing with such difficult manuscripts.

Abstract: We present a methodology to automatically obtain Pose Space Deformation (PSD) basis for rigged garments through deep learning. Classical approaches rely on Physically Based Simulations (PBS) to animate clothes. These are general solutions that, given a sufficiently fine-grained discretization of space and time, can achieve highly realistic results. However, they are computationally expensive and any scene modification prompts the need of re-simulation. Linear Blend Skinning (LBS) with PSD offers a lightweight alternative to PBS, though, it needs huge volumes of data to learn proper PSD. We propose using deep learning, formulated as an implicit PBS, to unsupervisedly learn realistic cloth Pose Space Deformations in a constrained scenario: dressed humans. Furthermore, we show it is possible to train these models in an amount of time comparable to a PBS of a few sequences. To the best of our knowledge, we are the first to propose a neural simulator for cloth.
While deep-based approaches in the domain are becoming a trend, these are data-hungry models. Moreover, authors often propose complex formulations to better learn wrinkles from PBS data. Supervised learning leads to physically inconsistent predictions that require collision solving to be used. Also, dependency on PBS data limits the scalability of these solutions, while their formulation hinders its applicability and compatibility. By proposing an unsupervised methodology to learn PSD for LBS models (3D animation standard), we overcome both of these drawbacks. Results obtained show cloth-consistency in the animated garments and meaningful pose-dependant folds and wrinkles. Our solution is extremely efficient, handles multiple layers of cloth, allows unsupervised outfit resizing and can be easily applied to any custom 3D avatar.

Abstract: The threat of 3D mask to face recognition systems is increasing serious, and has been widely concerned by researchers. To facilitate the study of the algorithms, a large-scale High-Fidelity Mask dataset, namely CASIA-SURF HiFiMask (briefly HiFiMask) has been collected. Specifically, it consists of total amount of 54,600 videos which are recorded from 75 subjects with 225 realistic masks under 7 new kinds of sensors. Based on this dataset and Protocol 3 which evaluates both the discrimination and generalization ability of the algorithm under the open set scenarios, we organized a 3D High-Fidelity Mask Face Presentation Attack Detection Challenge to boost the research of 3D mask based attack detection. It attracted more than 200 teams for the development phase with a total of 18 teams qualifying for the final round. All the results were verified and re-ran by the organizing team, and the results were used for the final ranking. This paper presents an overview of the challenge, including the introduction of the dataset used, the definition of the protocol, the calculation of the evaluation criteria, and the summary and publication of the competition results. Finally, we focus on introducing and analyzing the top ranked algorithms, the conclusion summary, and the research ideas for mask attack detection provided by this competition.