Abstract: Target encoding plays a central role when learning Convolutional Neural Networks. In this realm, one-hot encoding is the most prevalent strategy due to its simplicity. However, this so widespread encoding schema assumes a flat label space, thus ignoring rich relationships existing among labels that can be exploited during training. In large-scale datasets, data does not span the full label space, but instead lies in a low-dimensional output manifold. Following this observation, we embed the targets into a low-dimensional space, drastically improving convergence speed while preserving accuracy. Our contribution is two fold: (i) We show that random projections of the label space are a valid tool to find such lower dimensional embeddings, boosting dramatically convergence rates at zero computational cost; and (ii) we propose a normalized eigenrepresentation of the class manifold that encodes the targets with minimal information loss, improving the accuracy of random projections encoding while enjoying the same convergence rates. Experiments on CIFAR-100, CUB200-2011, Imagenet, and MIT Places demonstrate that the proposed approach drastically improves convergence speed while reaching very competitive accuracy rates.

Abstract: Wearable cameras can gather large amounts of image data that provide rich visual information about the daily activities of the wearer. Motivated by the large number of health applications that could be enabled by the automatic recognition of daily activities, such as lifestyle characterization for habit improvement, context-aware personal assistance and tele-rehabilitation services, we propose a system to classify 21 daily activities from photo-streams acquired by a wearable photo-camera. Our approach combines the advantages of a late fusion ensemble strategy relying on convolutional neural networks at image level with the ability of recurrent neural networks to account for the temporal evolution of high-level features in photo-streams without relying on event boundaries. The proposed batch-based approach achieved an overall accuracy of 89.85%, outperforming state-of-the-art end-to-end methodologies. These results were achieved on a dataset consists of 44,902 egocentric pictures from three persons captured during 26 days in average.

Abstract: Software as a Service (SaaS) is a cloud computing model in which a provider hosts applications in a server that customers use via internet. Since SaaS does not require to install applications on customers' own computers, it allows the use by multiple users of highly specialized software without extra expenses for hardware acquisition or licensing. A SaaS tailored for clinical needs not only would alleviate licensing costs, but also would facilitate easy access to new methods for diagnosis assistance. This paper presents a SaaS client-server architecture for Diagnosis as a Service (DaaS). The server is based on docker technology in order to allow execution of softwares implemented in different languages with the highest portability and scalability. The client is a content management system allowing the design of websites with multimedia content and interactive visualization of results allowing user editing. We explain a usage case that uses our DaaS as crowdsourcing platform in a multicentric pilot study carried out to evaluate the clinical benefits of a software for assessment of central airway obstruction.

Abstract: Automatic access control systems (ACS) based on the human biometrics or physical tokens are widely employed in public and private areas. Yet these systems, in their conventional forms, are restricted to active interaction from the users. In scenarios where users are not cooperating with the system, these systems are challenged. Failure in cooperation with the biometric systems might be intentional or because the users are incapable of handling the interaction procedure with the biometric system or simply forget to cooperate with it, due to for example, illness like dementia. This work introduces a challenging bimodal database, including face and hand information of the users when they approach a door to open it by its handle in a noncooperative context. We have defined two (an easy and a challenging) protocols on how to use the database. We have reported results on many baseline methods, including deep learning techniques as well as conventional methods on the database. The obtained results show the merit of the proposed database and the challenging nature of access control with non-cooperative users.

Abstract: We propose a novel attention mechanism to enhance Convolutional Neural Networks for fine-grained recognition. It learns to attend to lower-level feature activations without requiring part annotations and uses these activations to update and rectify the output likelihood distribution. In contrast to other approaches, the proposed mechanism is modular, architecture-independent and efficient both in terms of parameters and computation required. Experiments show that networks augmented with our approach systematically improve their classification accuracy and become more robust to clutter. As a result, Wide Residual Networks augmented with our proposal surpasses the state of the art classification accuracies in CIFAR-10, the Adience gender recognition task, Stanford dogs, and UEC Food-100.

Abstract: With the massive explosion of machine learning in our day-to-day life, incremental and adaptive learning has become a major topic, crucial to keep up-to-date and improve classification models and their corresponding feature extraction processes. This paper presents a categorized overview of incremental feature extraction based on linear subspace methods which aim at incorporating new information to the already acquired knowledge without accessing previous data. Specifically, this paper focuses on those linear dimensionality reduction methods with orthogonal matrix constraints based on global loss function, due to the extensive use of their batch approaches versus other linear alternatives. Thus, we cover the approaches derived from Principal Components Analysis, Linear Discriminative Analysis and Discriminative Common Vector methods. For each basic method, its incremental approaches are differentiated according to the subspace model and matrix decomposition involved in the updating process. Besides this categorization, several updating strategies are distinguished according to the amount of data used to update and to the fact of considering a static or dynamic number of classes. Moreover, the specific role of the size/dimension ratio in each method is considered. Finally, computational complexity, experimental setup and the accuracy rates according to published results are compiled and analyzed, and an empirical evaluation is done to compare the best approach of each kind.

Abstract: In this dissertation, we explore different ways to perform human action recognition in video clips. We focus on data efficiency, proposing new approaches that alleviate the need for laborious and time-consuming manual data annotation. In the first part of this dissertation, we start by analyzing previous state-of-the-art models, comparing their differences and similarities in order to pinpoint where their real strengths come from. Leveraging this information, we then proceed to boost the classification accuracy of shallow models to levels that rival deep neural networks. We introduce hybrid video classification architectures based on carefully designed unsupervised representations of handcrafted spatiotemporal features classified by supervised deep networks. We show in our experiments that our hybrid model combine the best of both worlds: it is data efficient (trained on 150 to 10,000 short clips) and yet improved significantly on the state of the art, including deep models trained on millions of manually labeled images and videos. In the second part of this research, we investigate the generation of synthetic training data for action recognition, as it has recently shown promising results for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation and other computer graphics techniques of modern game engines. We generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for “Procedural Human Action Videos”. It contains a total of 39,982 videos, with more than 1,000 examples for each action of 35 categories. Our approach is not limited to existing motion capture sequences, and we procedurally define 14 synthetic actions. We then introduce deep multi-task representation learning architectures to mix synthetic and real videos, even if the action categories differ. Our experiments on the UCF-101 and HMDB-51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance, outperforming fine-tuning state-of-the-art unsupervised generative models of videos.