Missions

This work is being carried out in the context of a joint laboratory bringing together research teams from IRISA and Ouest-France, France's leading regional daily newspaper. The aim is to challenge the latest computer vision techniques with the millions of photographs in Ouest-France's archives.

Activities

The Ouest-France archives contain over 35 million photographs, which are highly diverse in terms of period and content, but are very little used, mainly manually by documentalists.

Our first task is to understand the inner workings of various automatic analysis tools based on recent neural architectures (CNN, Transformers, etc.) specialized in image processing. This understanding is necessary to prove that the tools are suitable for processing photo archives in a context where many challenges arise, both of an applicative nature and in relation to the data and its properties. This proof will first be based on the analysis of a subset of the image collection, on which three tasks will be tested: the ability of modern tools to perform image classification, fine recognition of objects or details, and recognition of people and instances.
Challenges include learning classes from few examples and the dynamicity of this learning, the ability to identify all images of the same subject (for example, a place, a painting, or a statue) despite significant variations in light, framing, or other shooting conditions, but also despite changes in the appearance of the subjects themselves over time. Overall, this work will create links within vast collections of images, facilitating their exploration and structuring by curators. Feedback from documentalists and journalists will be central, as they will be the users of these technologies. A detailed analysis of computational requirements will be carried out, as computing resources are limited.

This proof-of-concept will then be used to guide analysis of the difficulties involved in scaling up, and the costs of learning and querying. This will be followed by work on adapting the learning and inference processes to the impossible knowledge of all the services and innovative downstream tasks that will exploit this visual corpus (in addition to tasks such as de-duplication, recommendation, similarity with diversity, dynamic grouping of image series according to various semantic criteria: same places, same characters, ...), automatic generation of descriptions or keywords, visualization of these images. Fundamentally, the work outlined here will tackle difficult scientific questions of image representation, metric learning, variety searches, continuous learning, data augmentation, model adaptation and multidimensional indexing. The aim is to identify the weaknesses of current techniques and propose improvements for applications in real-life contexts. An extension of these analyses will be to assess the capabilities of multimodal models such as CLIP and its descendants.

An occasional presence on the premises of Ouest-France, France's largest daily newspaper, located in Rennes, will be set up. Ouest-France has an ever-growing archive, where fast and accurate retrieval of instances and the creation of relevant navigation links are essential for journalists and documentalists. Absolutely unique datasets exist here, differing greatly from those used in all academic studies, which are often artificial and simplistic. It is certain that the best modern instance retrieval systems will be seriously put to the test in contexts of use with real images, offering opportunities for innovative research. A close partnership exists between Linkmedia and Ouest-France. The presence at Ouest-France will facilitate work with the archives and discussions with its users.

Skills

The candidate must :
- have a thesis in computer science or an engineering degree (or equivalent Bac+5) in computer science with a specialization in machine learning
- solid programming skills in Python, a high level of programming experience, and familiarity with modern software engineering techniques facilitating the reproducibility of results
- proficiency in English (B2) and French (A2) to analyze data and communicate with partners
- experience/specialization in computer vision or, failing that, in artificial learning and neural networks is essential
- be autonomous, have a rigorous scientific approach, be able to make proposals and take initiatives
- be motivated by a desire to understand how existing techniques work, in order to define their limits

Work Context

The position is located on the Beaulieu campus in Rennes.
Travel to Ouest-France's offices may be required.
Full-time position, 38h30 weekly with RTT.
Telecommuting possible.

About the laboratory
=============
www.irisa.fr
IRISA is one of France's largest research laboratories (over 850 staff) in the field of computer science and information technology. Structured into seven scientific departments, IRISA is a laboratory of excellence whose scientific priorities are bioinformatics, systems security, new software architectures, virtual reality, Big Data analysis and artificial intelligence. With its sights set on the future of information technology and its international outlook, IRISA is at the very heart of society's digital transition and of innovation in cybersecurity, health, the environment and ecology, transport, robotics, energy, culture and artificial intelligence.

Presentation of CNRS as an employer: https://www.cnrs.fr/fr/le-cnrs
Presentation of IRISA as laboratory of assignment: https://www.irisa.fr/umr-6074

The position is located in a sector covered by the protection of scientific and technical potential (PPST), and may require, in accordance with regulations, that your arrival be authorized by the competent MESR authority.

Additional Information

This work is carried out in the context of the SYNAPSES joint laboratory.