SNSF | P3 Research Database

Object Detection; Pose Estimation; Computer Vision; Object Recognition; Feature Point Recognition; 3D Pose

English title	View Sets for 3-D Object Detection and Recognition
Applicant	Lepetit Vincent
Number	116195
Funding scheme	Project funding (Div. I-III)
Research institution	Laboratoire de vision par ordinateur EPFL - IC - ISIM - CVLAB
Institution of higher education	EPF Lausanne - EPFL
Main discipline	Information Technology
Start/End	01.07.2007 - 30.06.2009
Approved amount	103'433.00

Lead

Lay summary
During the first year of this project, we have shown how to achieve robust real-time 3D object detection even where no a priori 3D model is available. The key ingredient of the approach we developed is what we refer to as feature harvesting, in which we robustly learn both geometry and appearance of the target object from a video sequence. We now propose to extend feature harvesting to the recognition of all objects that belong to a class, such as cars, faces, or mugs. To this end, we will first use our current algorithm to learn individual models using several video sequences of specific instances. We will then establish correspondences across instances and use them to register the models. Our preliminary experiments show that this should be relatively easy because inter-object appearance variations is not systematically large, which should allow us to select appropriate image pairs for matching purposes. This, in turn, will let us cluster image patches that correspond to the same part of objects of the same category under similar viewpoints and learn the corresponding appearance models. The resulting generic models will store the cluster labels distributions over the object. To avoid mis-detections due to keypoint extraction failures, we will forgo feature extraction and perform dense labeling. Because an exhaustive search through the possible modelisations would then be prohibitively expensive, we will develop a run-time mechanism to quickly select the best ones. This will result in a system able to recognize classes of objects from arbitrary viewpoints, which is beyond the current state-of-the-art: Even though good results have been demonstrated for narrow ranges of viewpoint, almost none of the existing techniques can handle arbitrary ones. We therefore intend to address an open and fundamental Computer Vision problem, the automated detection and pose estimation of 3--D objects in real-world conditions. Since humans can do this very easily, whereas computers still cannot, this is an important scientific issue. Furthermore, the practical real-time solution we propose to develop will have many potential applications in fields ranging from Robotics and Human-Machine Interfaces to Computer Graphics.