en

Post-doctoral position in computer vision (M/F)

Post-doctoral position in computer vision (M/F)

France 28 Feb 2021
Ifremer - French Research Institute for Exploitation of the Sea

Ifremer - French Research Institute for Exploitation of the Sea

State University (France), Browse similar opportunities

OPPORTUNITY DETAILS

Total reward
0 $
State University
Area
Host Country
Deadline
28 Feb 2021
Study level
Opportunity type
Specialities
Opportunity funding
Not funding
Eligible Countries
This opportunity is destined for all countries
Eligible Region
All Regions

Reference: PV-2021-874/1

Department/Office:

Duration of contract:18 mois

Start date:

Deadline for applications:28/02/2021

The Institute and the recruiting department

A pioneer in ocean science, IFREMER’s cutting-edge research is grounded in sustainable development and open science. Our vision is to advance science, expertise and innovation to:

- Protect and restore the ocean

- Sustainably use marine resources to benefit society

- Create and share ocean data, information & knowledge.

With more than 1,500 personnel spread along the French coastline in more than 20 sites, the Institute explores the 3 great oceans: the Indian, Atlantic and Pacific oceans. A leader in ocean science, IFREMER is managing the French Oceanographic Fleet and its dedicated scientists create ground-breaking technology to push the boundaries of ocean exploration and knowledge, from the abyss to the atmosphere-ocean interface.

Well-established in the international scientific community, our scientists, engineers and technicians are committed to advance knowledge about our planet’s last unexplored frontiers. They provide the science we need for informed decision-making and public policy and they transfer this knowledge and technology to businesses to fulfill public and private needs. Core to our mission is also to strengthen public awareness about the importance of understanding the ocean and its resources, and empowering future generations of leaders through education and outreach national campaigns.

Founded in 1984, IFREMER is a French public organization and its budget approximates 240 million euros. It is operating under the joint authority of the French Ministry for Higher Education, Research and Innovation, the French Ministry for the Ecological and Solidary Transition, and the French Ministry of Agriculture and Food.

General areas of responsibility

Title :

Boosted Annotations of 3D models of deep-sea hydrothermal mounds by unsupervised domain adaptation  – ABYSSES

Keywords: Habitat mapping, hydrothermal vents, underwater imagery, machine learning, transfer learning

Abstract

The use of underwater imagery has become an essential tool in the assessment of biodiversity and the ecological monitoring of marine ecosystems [[1] Durden JM et al. (2016) Perspectives in visual imaging for marine biology and ecology: From acquisition to understanding. Oceanogr. Mar. Biol. 54:1-72]. Among these approaches, the 3D reconstruction of the seabed allows a detailed characterization, over large areas (ie km) of the biological (animal communities, habitat, species) and environmental (substrate, slope, roughness, topography) attributes of benthic ecosystems [[2] Robert K et al. (2017) New approaches to high-resolution mapping of marine vertical structures. Sci Rep 7:9005; [3] Gerdes K et al. (2019) Detailed mapping of hydrothermal vent fauna: A 3d reconstruction approach based on video imagery. Front. Mar. Sci. 6: 96].

Hydrothermal vents, located along mid-ocean ridges, are characterized by complex uneven topographies (i.e. escarpments, chimneys, crevasses) which play an important role in the distribution of biodiversity [[4] Girard F et al. Currents and topography drive assemblage distribution on an active hydrothermal edifice. Prog. Oceanogr. (in review)]. They are therefore particularly well suited to these 3D approaches. But the annotation of such models is time-consuming and requires more analytical power than the human capacity of the laboratory. Deep Learning methods now allow for the development of automatic processing capacities of large databases. Their performance in identification enables the provision of tools for processing complex images of the marine environment, a field reserved for rare expertise. But a convolutional network requires a long training phase from numerous annotated images, and current tools and methods do not guarantee sufficiently qualitative and quantitative annotation.

To address this problem, we assume that a part of the visual characteristics of hydrothermal ecosystems are shared among other visual domains for which the availability of annotated data is more substantial. Indeed, while the availability of annotated data was crucial for the development of deep neural network architectures for specific domains, nowadays, the proliferation of annotated ground truth data in multiple and diverse domains allows to compensate the lack of data for a new domain through its affinity with existing ones. Research in the domain known as transfer learning [[5] S. J. Pan et al. (2010), A Survey on Transfer Learning, in IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp. 1345-1359, 2010] concerns the family of methods that allow to emerge such affinities in an unsupervised manner. In this context, we are particularly interested in the development of methods belonging to the family of “transductive transfer learning” or “unsupervised domain adaptation” [[6] G. Kang et al. (2019)], Contrastive Adaptation Network for Unsupervised Domain Adaptation, IEEE Int. Conf. on Computer Vision and Pattern Recognition]. The objective of our study will be thus to establish the affinity of hydrothermal ecosystems in correspondence with terrestrial surfaces, mapped by mobile or aerial vehicles. Minimising the discrepancy between these domains is therefore expected to reduce the need for fully supervised annotation of hydrothermal images, by prototyping a weakly-supervised annotation tool.

This project, based on the complex example of hydrothermal ecosystems, is a first step towards the implementation of new digital tools in order to accelerate our ability to map, at high resolution, and over large spatial areas, the biological, environmental and topography of the seabed. The development of such tools will increase our ability to explore deep benthic ecosystems and therefore increase the acquisition of new knowledge in these environments.

Required Knowledge, skills, and characteristics

Required skills: PhD in Computer science with competences in artificial intelligence, computer vision and machine learning

Specific working conditions

Internal and external collaborations

In the lab, the candidate will interact with Marjolaine Matabos, researcher in deep-sea benthic ecology, and Catherine Borremans, engineer in imagery. He/she will also collaborate with Aurélien Arnaubec (Ifremer Toulon) regarding technical aspects.

The project is conducted in collaboration with colleagues at the CERV (European Center on Virtual Reality), Panagiotis Papadakis (IMTA) and Cédric Buche (ENIB), in charge of the methodological development within the project. He/she will work and exchange on a regular basis with them.

How to apply for this position

Deadline for applications: 28/02/2021

All applications are processed exclusively via our website. Interested candidates can apply by clicking the “Apply” button. 

Other organizations


Choose your study destination


Choose the country you wish to travel to study for free, work or volunteer

Please find also