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: 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: Developing arbitrary symbol recognition systems is a challenging endeavour. Even using content-agnostic architectures such as few-shot models, performance can be substantially improved by providing a number of well-annotated examples into training. In some contexts, transcripts of the symbols are available without any position information associated to them, which enables using line-level recognition architectures. A way of providing this position information to detection-based architectures is finding systems that can align the input symbols with the transcription. In this paper we discuss some symbol alignment techniques that are suitable for low-data scenarios and provide an insight on their perceived strengths and weaknesses. In particular, we study the usage of Connectionist Temporal Classification models, Attention-Based Sequence to Sequence models and we compare them with the results obtained on a few-shot recognition system.

Abstract: CoRR abs/1607.08811
Nowadays, we can find several diseases related to the unhealthy diet habits of the population, such as diabetes, obesity, anemia, bulimia and anorexia. In many cases, these diseases are related to the food consumption of people. Mediterranean diet is scientifically known as a healthy diet that helps to prevent many metabolic diseases. In particular, our work focuses on the recognition of Mediterranean food and dishes. The development of this methodology would allow to analise the daily habits of users with wearable cameras, within the topic of lifelogging. By using automatic mechanisms we could build an objective tool for the analysis of the patient’s behavior, allowing specialists to discover unhealthy food patterns and understand the user’s lifestyle.
With the aim to automatically recognize a complete diet, we introduce a challenging multi-labeled dataset related to Mediter-ranean diet called FoodCAT. The first type of label provided consists of 115 food classes with an average of 400 images per dish, and the second one consists of 12 food categories with an average of 3800 pictures per class. This dataset will serve as a basis for the development of automatic diet recognition. In this context, deep learning and more specifically, Convolutional Neural Networks (CNNs), currently are state-of-the-art methods for automatic food recognition. In our work, we compare several architectures for image classification, with the purpose of diet recognition. Applying the best model for recognising food categories, we achieve a top-1 accuracy of 72.29%, and top-5 of 97.07%. In a complete diet recognition of dishes from Mediterranean diet, enlarged with the Food-101 dataset for international dishes recognition, we achieve a top-1 accuracy of 68.07%, and top-5 of 89.53%, for a total of 115+101 food classes.

Abstract: Local image features are often used to efficiently represent image content. The limited number of types of features that a local feature extractor responds to might be insufficient to provide a robust image representation. To overcome this limitation, we propose a context-aware feature extraction formulated under an information theoretic framework. The algorithm does not respond to a specific type of features; the idea is to retrieve complementary features which are relevant within the image context. We empirically validate the method by investigating the repeatability, the completeness, and the complementarity of context-aware features on standard benchmarks. In a comparison with strictly local features, we show that our context-aware features produce more robust image representations. Furthermore, we study the complementarity between strictly local features and context-aware ones to produce an even more robust representation.

Abstract: Issues such as content identification, document and image security, audience measurement, ownership and copyright among others can be settled by the use of digital watermarking. Many recent video watermarking methods show drops in visual quality of the sequences. The present work addresses the aforementioned issue by introducing a robust and imperceptible non-blind color video frame watermarking algorithm. The method divides frames into moving and non-moving parts. The non-moving part of each color channel is processed separately using a block-based watermarking scheme. Blocks with an entropy lower than the average entropy of all blocks are subject to a further process for embedding the watermark image. Finally a watermarked frame is generated by adding moving parts to it. Several signal processing attacks are applied to each watermarked frame in order to perform experiments and are compared with some recent algorithms. Experimental results show that the proposed scheme is imperceptible and robust against common signal processing attacks.