Marie-Odile Berger

Senior Researcher

INRIA Lorraine

ISA team

Open Positions : (Post-doc)

Post-doctorat: Structure from motion in multi-planar environments for Augmented Reality Applications. (pdf) Apply

Research domains:

Medical Imaging:

(René Anxionnat, Marie-Odile Berger, Erwan Kerrien, Jean-Paul Lethor, Goetz Winterfeldt)

3D reconstruction of the beating left ventricle from echo-cardiographic images using a rotating probe (See Méthod and demos) .
Interventional Neuroradiology:
- Fully automatic 2D/3D registration of X-ray angiography images
- Use of multimodality to perform arterio-veinous malformation delineation

Augmented reality: Adding virtual objects to video sequences.

(Michael Aron, Marie-Odile Berger, Vincent Lepetit, Gilles Simon, Flavio Vigueras)

Our objective is to add virtual objects to real video sequences in such a manner as to appear part of the 3D world. In order to make augmented reality systems effective, the computer generated objects must be combined seemlessly so that the virtual objects align well with the real one. The virtual objects must be rendered from the same sequence of viewpoints as the original footage. Realistic merging of virtual and real objects also requires that objects behave in a physical plausible amnner in the environment: they can be occluded by objects in the scene, they are shadowed by other objects. In the Isa team, we have developed vision based methods appropriate for image composition both for viewpoint computation and occlusion resolution.

Viewpoint Computation:

We first developed a robust and efficient method for temporal registration. This method is fully automatic. This method takes advantage of 3D knowledge that are available on the scene as well as 2D/2D point correspondences over time ( Demos ). We have recently extended this method to the case of cameras with varying focal length. The underlying assumption is that the zoom and the camera pose do not change at the same time (Demos) .

Handling Occlusion between real and virtual objects:

As the virtual object can be occluded by real objects in the scene, we have to determine the visible part of the virtual object. Theoretically, resolving occlusions amounts to compare the depth of the scene to the depth of the virtual object. However, computing dense and accurate depth map is difficult. In addition, the interframe motion is not known in AR applications but must be computed. This can result in possibly large reconstruction errors which makes the 3D reconstruction untractable.

To overcome this problem, we propose a semi-interactive approach based on key-views to make easier the reconstruction process. Once the occluding object has been outlined in the key-views, our system detects automatically the occluding objects in the in-between views using local 3D reconstructin and region based tracking. This allows us to get an accurate delineation of the occluding objects (method et demo) .

Real time camera tracking for scenes containing planar structures:

Within the europeen project ARIS, we have developped a markerless system for real time camera tracking in multi-planar environnements. The main contributions of our system to the current state of the art are twofold:

selection model is used to reduce fluctuations in the parameters that are commonly encountered especially when the camera motion is small. Using model selection we get a better visual impression of the augmented scene. In addition, we prove that the use of model selection improves the accuracy of the viewpoint and reduces noticeably the drift problem

As pure vision-based tracking methods cannot usually keep up with fast or abrupt user movements, a hybrid approach which combines the accuracy of vision-based methods and the robustness of inertial tracking systems has been designed. This allows us to improve dramatically the robustness of the system.
See methods and demos