Pattern Recognition in Two National Labs
I recently attended a research review at France’s INRIA. For the past 7 years, I have attended reviews for another government-sponsored lab in the US, NIST. Although these labs are different in their charter and organization, both sponsor work in the area of pattern recognition (among many other topics), so I thought the pattern recognition community would be interested to learn the processes and procedures by which science and technology are fostered at these two organizations.
First, I should note that all of the material in this article is publicly available—no confidential material from reviews is discussed here. These labs, being publicly funded, make efforts to publicize their efforts. Second, this is not a comparative article since these two labs have different objectives.
Pattern Recognition at INRIA
INRIA, the French Institut National du Recherche en Informatique et Automatique (National Institute for Research in Computer Science and Control), was founded in 1967. The Institute is under the dual authority of the Ministries of Research and Industry. In numbers, there are 6 research units that are geographically dispersed across France, a staff of 3500 including 2700 scientists, and 124 project-teams. The INRIA budget in 2004 was 123 million Euros. More than half the projects involve collaborations with CNRS (Centre Nationale de la Recherche Scientifique), universities, and industrial partners.
As part of its research objective, INRIA actively fosters international collaborations. For instance, there are 9 associate teams, mainly now in the US, however increasingly in Asian countries as well. Training and education are also high priorities with 950 doctoral candidates, 150 post-docs, and 300 trainees, all involved on INRIA projects. In addition to research, technology transfer is also a stated strategic objective.
It is interesting to note that, in the organizational hierarchy between INRIA administration and the research projects, there are no other levels; there are no divisions, labs, or departments. Instead, there are just the 124 project-teams that are each centered at one of the 6 INRIA locations, although collaboration may extend beyond the location and indeed internationally. Teams consist of about 15-25 members. These members may be full-time INRIA employees, but also may be professors, post-docs, graduate students, visiting researchers, and industry collaborators.
INRIA has 5 themes of its research, into which each project-team is categorized. The communicating systems theme includes work in networking, telecom, and mobility. The cognitive systems theme includes work in machine learning, graphics, computer vision, and multimedia data. The symbolic systems theme includes work in software languages, algorithms, and reliability. The numerical systems theme includes work in automatic control, robotics, signal processing, and scientific computing. The biological systems theme includes work in modeling and simulation for medicine and biology.
Many of the Cognitive Systems projects either employ pattern recognition techniques or are directly pursuing research on pattern recognition methods. Some projects that will be of interest to IAPR Newsletter readers are briefly described here.
The earth monitoring group, called ARIANA, is performing work involving analysis of aerial and satellite images. Major objectives in this work include segmentation and interpolation of land areas, roads, rivers, and buildings. There are many conditions particular to the capture mode that makes this work distinct and difficult. For instance, occlusion is a difficult condition, whether caused by clouds or trees, and the rules that are used for continuity of occluded parts of rivers and roads can lead to errors. Resolution is becoming higher, but is still not high enough to fully recognize many ground items except by the texture of segmented areas of these items (such as trees). Finally, the size of images and the volume of data that are produced by the satellites are so high that it is impossible to fully analyze all images, so fast image search and indexing techniques are also important.
The TEMICS group is involved in coding, communications, and watermarking of still and video images. A major goal of this group is scalable compression, that is a compression scheme that would enable video suppliers to transmit their material to devices of different size and bandwidth capabilities, all from a single compressed source, but each optimal in speed and quality to its destination.
Finally, QGAR, a graphics recognition project, is seeking to extend the bounds of symbol and line recognition. Besides facilitating image understanding and symbolic storage of the document, symbol recognition enables quick searches of document images. One can think of this as a mode of search in between textual and gray-scale picture search. When we observe the great strides made in OCR of machine text in the past decade, and the still early stages of automatic image indexing, one can understand that graphics recognition is still a challenging problem whose level of progress is somewhere between these two.
Pattern Recognition at NIST
NIST the National Institute of Standards and Technology was founded in 1901 to promote US innovation and industrial competitiveness by advancing measurement science, standards, and technology. It is under the US Commerce Department. NIST employs about 3000 scientists, technicians, and administrative staff. The operating budget is about $858 million. There are two NIST campuses, one in Maryland and the other in Colorado.
The names of the 7 labs go far to describe the work done at NIST: Building and Fire Research, Chemical Science and Technology, Electronics and Electrical Engineering, Information Technology, Manufacturing Engineering, Materials Science and Engineering, and Physics. The Information Technology Lab contains work that centers around computers and networks. Its divisions are: Mathematical and Computational Sciences, Advanced Networking Technologies, Computer Security, Information Access, Software Diagnosis and Conformance Testing, and Statistical Engineering. The Information Access Division contains a number of projects that have pattern recognition components. Some of these are described below.
One of the major efforts at this time is in biometrics. The US is leading the world in government adoption of biometrics. For instance, the US VISIT program, which has been put in place in the last two years, requires fingerprint and face images for visitors with visas to the US. As part of this effort, the NIST biometrics group has compiled massive databases of fingerprints (totaling about 100 million fingerprints for 16 million subjects) and faces for testing. This group is involved in work to specify fingerprint standards for storage and interoperability. They have performed fingerprint vendor technology evaluations. They have also created and run competitions on face and gait recognition. Fingerprint databases are freely available for development and testing and are being used by researchers and biometrics developers around the world.
Speech is another area in which NIST performs its unique work to drive that technology forward, to measure its state-of-the-art recognition rates, and to find the most promising algorithmic approaches. The procedure for doing this is typically the following: an evaluation is announced, participants are provided training data, participants submit results on test data, evaluations are performed, and a workshop is convened to present and discuss results. Current and upcoming evaluations include: automatic content extraction; classification of locales, events, activities, and relationships; machine translation; language recognition; machine translation; rich transcription; and speaker recognition.
The final project described here is TREC, (Text Retrieval Conferences). This series of conferences has taken place yearly since 1992 and forms focal points of the natural language and text retrieval communities. The goal of the conference series is to encourage research in information retrieval by providing a large test collection, uniform scoring procedures, and a forum for organizations interested in comparing their results. Some tasks involved here are: topic detection and tracking, advanced questioning and answering, and automatic summarization. A recent spin-off from the textual focus of TREC is TRECVID, whose objective is to perform automatic segmentation, indexing, and content-based retrieval of digital video.
For more information on these labs and a complete list of their projects, see their web sites.