Sites Inria

English version

Equipe de recherche TEXMEX

Rapports d'activité

Overall Objectives

With the success of sites like Youtube or DailyMotion, with the development of the Digital Terrestrial TV, it is now obvious that the digital videos have invaded our usual information channels like the web. While such new documents are now available in huge quantities, using them remains difficult. Beyond the storage problem, they are not easy to manipulate, browse, describe, search, summarize, visualize as soon as the simple scenario “1. search the title by keywords 2. watch the complete document” does not fulfill the user's needs anymore. That is, in most cases.

Most usages are linked with the key concept of repurposing. Videos are a raw material that each user recombines in a new way, to offer new views of the content, to adapt it to new devices (ranging from HD TV sets to mobile phones), to mix it with other videos, to answer information queries... Somehow, each use of a video gives raise to a new short-lived document that exists only while it is viewed. Achieving this repurposing process implies the ability to manipulate videos extracts as easily as words in a text.

Many applications exist in both professional and domestic areas. On the professional side, these applications include transforming a TV broadcast program into a web site, a DVD or a mobile phone service, switching from a traditional TV program to an interactive one, better exploiting TV and video archives, constructing new video services (video on demand, video edition, etc). On the domestic side, video summarizing can be of great help, as can a better management of the videos locally recorded, or simple tools to face the exponential number of TV channels available that increase the quantity of interesting documents available, overall increasing but make them really hard to find.

In order to face such new application needs, we propose a multi-field work, gathering in a single team specialists that are able to deal with the various media and aspects of large video collections: image, video, text, sound and speech, but also data analysis, indexing, machine learning... The main goal of this work is to segment, structure, describe, or de-linearize the multimedia content in order to be able to recombine or re-use that content in new conditions. The focus on the document analysis aspect of the problem is an explicit choice since it is the first mandatory step of any subsequent application, but using the descriptions obtained by the processing tools we develop is also an important goal of our activity.

To illustrate our research project in one short sentence, we would like our computers to be able to watch TV and use what has been watched and understood in new innovative services. The main challenges to address in order to reach that goal are: the size of the documents and of the document collections to be processed, the necessity to process jointly several media and to obtain a high level of semantics, the variety of contents, of contexts, of needs and usages, linked to the difficulty to manage such documents on a traditional interface.

Our own research is organized in three directions: 1- developing advanced algorithms of data analysis, description and indexing, 2- searching new techniques for linguistic information acquisition and use, 3- building new processing tools for audiovisual documents.

Advanced algorithms of data analysis, description and indexing

Processing multimedia documents produces most of the time lots of descriptive metadata. These metadata can take many different aspects ranging from a simple label issued from a limited list, to high dimensional vectors or matrices of any kind; they can be numeric or symbolic, exact, approximate or noisy. As examples, image descriptors are usually vectors whose dimension can vary between 2 and 900, while text descriptors are vectors of much higher dimension, up to 100,000 but that are very sparse. Real size collections of documents can produce sets of billions of such vectors.

Most of the operations to be achieved on the documents are in fact translated in terms of operations on their metadata, which appear as key objects to be manipulated. Although their nature is much simpler than the data used to compute them, these metadata require specific tools and algorithms to cope with their particular structure and volume. Our work concerns mainly three domains:

data analysis techniques, possibly coupled to data visualization techniques, to study the structure of large sets of metadata, with applications to classical problems like data classification, clustering, sampling, or modeling,

advanced data indexing techniques in order to speed-up the manipulation of these metadata for retrieval or query answering problems,

description of compressed, watermarked or attacked data.

New techniques for linguistic information acquisition and use

Natural languages are a privileged way to carry high level semantic information. Used in speech from an audio track, in textual format or overlaid in images or videos, alone or associated with images, graphics or tables, organized linearly or with hyperlinks, expressed in English, French, or Chinese, this linguistic information may take many different forms, but always exhibits a common basic structure: it is composed of sequences of words. Building techniques that preserve the subtle links existing between these words, their representations with letters or other symbols and the semantics they carry is a difficult challenge.

As an example, actual search engines work at the representation level (they search sequences of letters), and do not consider the meaning of the searched words. Therefore, they do not use the fact that “bike” and “bicycle” represent a single concept while “bank” has at least two different meanings (a river bank and a financial institution).

Extracting high level information is the goal of our work. First, acquisition techniques that allow us to associate pieces of semantics with words, to create links between words are still an active field of research. Once this linguistic information is available, its use raises new issues. For example, in search engines, new pieces of information can be stored and the representation of the data can be improved in order to increase the quality of the results.

New processing tools for audiovisual documents

One of the main characteristic of audiovisual documents is their temporal dimension. As a consequence, they cannot be watched or listened to globally, but only by a linear process that takes some time. On the processing side, these documents often mix several media (image track, sound track, some text) that should be all taken into account to understand the meaning and the structure of the document. They can also have an endless stream structure with no clear temporal boundaries, like on most TV or radio channels. Therefore, there is an important need to segment and structure them, at various scales, before describing the pieces that are obtained.

Our work is organized in three directions. Segmenting and structuring long TV streams (up to several weeks, 24 hours a day) is a first goal that allows to extract program and non program segments in these streams. These programs can then be structured at a finer level. Finally, once the structure is extracted, we use the linguistic information to describe and characterize the various segments. In all this work, the interaction between the various media is a constant source of difficulty, but also of inspiration.

Suivez Inria