- Cet évènement est passé
FIL Talk – January, 19
19 janvier 2023 à 15 h 00 min - 16 h 00 min
We have a pleasure to invite you to the FIL seminar.
When: January 19, 15:00 – 16:00.
Where: Campus de la Doua – Bâtiment Hedy Lamarr- Département Télécommunications
Amphithéâtre Claude Chappe
6 Av. des Arts, 69100 Villeurbanne
Talk by: Johanna Delanoy
Title: Perception and manipulation of material appearance
Keywords: computer graphics, material appearance, perception
Material appearance is key to the way we visually perceive the world around us: the fact that an object appears metallic, glossy or matte strongly impact how we perceive such objects. However, it is not yet fully understood how we perceive such material properties and how to manipulate them in images. In this presentation, I will address these two aspects. First, I will quickly present a study in which we studied the perception of material appearance by using the implicit knowledge from painters. To achieve this goal, we directly compared the perception of material properties in paintings and in renderings, by collecting professional realistic paintings of rendered materials. Then, I will focus on an image-based editing method that allows to modify the material appearance of an object by increasing or decreasing high-level perceptual attributes (such as « Metallic » or « Glossy »). Our method uses a single image as input and does not require any 3D representation of the scene. We used deep generative networks, those representations were shown to correlate with human perception, that we trained with human perceptual data. This combination of perceptual data and deep learning allows us to modify images in a way that is perceptually plausible.
Johanna Delanoy is interested in using digital tools to help artists create or explore the creation space based on perceptual knowledge. She likes to study how artists work with traditional techniques and how we perceive what they create in order to propose better solutions to assist creative tasks.
Talk by: Stéphanie Jehan-Besson
Title: Optimization of a shape metric based on information theory for segmentation evaluation and fusion in 2D or 3D medical image analysis
Keywords: segmentation fusion and evaluation, computation of mean shapes, information theory, shape gradients, medical image analysis in 2D or 3D
In the field of delineation of 2D or 3D complex shapes in medical imaging, and especially due to the development of multimodal and multiparametric image acquisition devices, the combination of segmentation results from different sources is interesting. It is also essential to accurately assess the variability between delineation experts or results obtained with different algorithms and different parameterization. In this work, we propose to estimate what we call a mutual shape and define it as the optimum of a statistical criterion based on information theory. This criterion is justified by using the similarities between the information theory quantities and the area measures, which leads us to interpret the mutual shape as a shape metric belonging to the Fréchet family. The information theory quantities are estimated using probability density functions in a continuous variational framework. The mutual shape is then computed using shape optimization tools through the calculation of shape gradients. We provide synthetic and real examples in 2D or 3D to demonstrate the applicability of our framework for evaluation and fusion. These examples allow us to highlight the difference between the mutual shape and a classical mean shape. We also compare the mutual shape to the well known STAPLE method and to the minimization of a simple symmetric difference. We provide experimental results in medical imaging for the combination and evaluation of segmentation results of cardiac structures in MRI or echocardiography. Different perspectives will be mentioned, such as multimodal data fusion in MRI or the integration of expert variability assessment in deep learning methods.
After a PhD in the I3S Laboratory on the topic of deformable models based on shape gradients and domain optimization for image and video segmentation, Stéphanie Jehan-Besson became an associate professor at the engineering school of Caen (ENSICAEN). In 2008, she obtained a permanent position as a CNRS researcher in the field of image and video processing, first at the LIMOS Laboratory, then at the GREYC Laboratory in Caen. In 2021, she joined the Curie Institute as a visiting researcher in the LITO Laboratory (INSERM U1288). In January 2022, she moved to the CREATIS laboratory as a permanent CNRS researcher in the MYRIAD team. Her research interests concern the design and optimization of continuous statistical criteria for various applications (segmentation of complex structures in 2D or 3D images, motion estimation and tracking in video sequences, filtering based on total variation, evaluation or fusion of segmentation methods). While continuing this work and applications to various medical image challenges, she is exploring deep learning methods for these applications.