Section: New Results

Multimedia content description and structuring

Hierarchical topic structuring

Participants : Guillaume Gravier, Pascale Sébillot.

In [37], we investigated the potential of a topical structure of text-like data that we recently proposed  [55] in the context of summarization and anchor detection in video hyperlinking. This structure is produced by an algorithm that exploits temporal distributions of words through word burst analysis to generate a hierarchy of topically focused fragments. The obtained hierarchy aims at filtering out non-critical content, retaining only the salient information at various levels of detail. For the tasks we choose to evaluate the structure on, the lost of important information is highly damaging. We show that the structure can actually improve the results of summarization or at least maintain state-of-the-art results, while for anchor detection it leads us to the best precision in the context of the Search and Anchoring in Video Archives task at MediaEval. The experiments were carried on written text and a more challenging corpus containing automatic transcripts of TV shows.

Multimedia-inspired descriptors for time series classification

Participant : Simon Malinowski.

The SIFT framework has shown to be effective in the image classification context. Recently, we designed a bag-of-words approach based on an adaptation of this framework to time series classification. It relies on two steps: SIFT-based features are first extracted and quantized into words; histograms of occurrences of each word are then fed into a classifier. In [38], we investigated techniques to improve the performance of bag-of-temporal-SIFT-words: dense extraction of keypoints and different normalizations of Bag-of-Words histograms. Extensive experiments have shown that our method significantly outperforms nearly all tested standalone baseline classifiers on publicly available UCR datasets. In [23], we also investigate the use of convolutional neural networks (CNN) for time series classification. Such networks have been widely used in many domains like computer vision and speech recognition, but only a little for time series classification. We have designed a convolutional neural network that consists of two convolutional layers. One drawback with CNN is that they need a lot of training data to be efficient. We propose two ways to circumvent this problem: designing data-augmentation techniques and learning the network in a semi-supervised way using training time series from different datasets. These techniques are experimentally evaluated on a benchmark of time series datasets.

Early time series classification

Participant : Simon Malinowski.

In time series classification, two antagonist notions are at stake. On the one hand, in most cases, the sooner the time series is classified , the higher the reward. On the other hand, an early classification is more likely to be erroneous. Most of the early classification methods have been designed to take a decision as soon as a sufficient level of reliability is reached. However, in many applications, delaying the decision with no guarantee that the reliability threshold will be met in the future can be costly. Recently, a framework dedicated to optimizing the trade-off between classification accuracy and the cost of delaying the decision was proposed, together with an algorithm that decides online the optimal time instant to classify an incoming time series. On top of this framework , we have built in [29] two different early classification algorithms that optimize the trade-off between decision accuracy and the cost of delaying the decision. These algorithms are non-myopic in the sense that, even when classification is delayed, they can provide an estimate of when the optimal classification time is likely to occur. Our experiments on real datasets demonstrate that the proposed approaches are more robust than existing methods.