Offre en lien avec l’Action/le Réseau : – — –/Doctorants

Laboratoire/Entreprise : IRISA/UBS
Durée : 6 mois
Contact : minh-tan.pham@irisa.fr
Date limite de publication : 2024-04-01

Contexte :
Detailed topic at: http://www-obelix.irisa.fr/files/2023/11/2023_master_topic_MTL.pdf

In recent years, deep neural networks have been successfully adopted in almost every application
domains of computer vision, including remote sensing for earth observation. The vast number
of remote sensing images captured from frequent satellite passes or aerial acquisition, however,
are not readily usable to train deep networks developed for generic vision problems due to the
lack of task-specific annotations and possible domain gaps.
On the other hand, the individual development efforts of various research groups for their
particular problems result in cluttered annotations and modalities: each dataset is typically
annotated for a few tasks while many tasks may be related to one another and could be jointly
learned to leverage complementary information and improve their performance. Coupling solving
different but related tasks, or well-known in the ML community as multi-task learning, has also
gained increasing attention in the remote sensing community. As multi-task learning aims to
predict different targets from the same inputs, it typically requires annotations of all the target
tasks for each input example to learn the interrelationship at the shared encoder by optimizing
all tasks at the same time.
Obtaining extra annotations to maintain multi-task datasets, however, add extra burden
to the development process. Recently, it has been shown in the vision community that that
multi-task learning could be beneficial even when the tasks are partially annotated [2]. Training
a network for multiple task while the training examples are annotated for a single task can
improve the performance of both tasks. Such discovery could be of interest to explore for the
benefit of remote sensing community.

Sujet :
This project is aimed to research the combination of different datasets annotated for different
tasks which may follow different statistical distributions to benefit and improve performance of
one another. To that end, we will focus on the object detection, i.e. bounding boxes prediction,
and semantic segmentation tasks, which are closely related yet not trivial to combine due to differences in spatial structure and information granularity: object detection predicts bounding-box
coordinates at object instance level while semantic segmentation provides per-pixel predictions
of category including amorphous regions. A general scheme is shown in Figure 1. Another
challenge of the project is to bridge possible domain gaps between the participating datasets
with possible approaches including generative models (GANs, diffusion models, etc.)

Profil du candidat :
Student in computer science and/or machine learning and/or signal & image processing;

Formation et compétences requises :
Python programming and familiarity with deep learning framework (Pytorch/Tensorflow);

Adresse d’emploi :
IRISA (UMR 6074) is located in the UBS (Université Bretagne Sud), campus de Tohannic, Vannes 56000, France.

Document attaché : 202311201646_2023_master_topic_MTL.pdf