Lifting 2D Object Detections to 3D: A Geometric Approach in Multiple Views

We present two new methods based on Interval Analysis and Computational Geometry for estimating the 3D occupancy and position of objects from image sequences. Given a calibrated set of images, the proposed frameworks first detect objects using off-the-shelf object detectors and then match bounding boxes in multiple views. The 2D semantic information given by the bounding boxes are used to efficiently recover 3D object position and occupancy using solely geometrical constraints in multiple views. We also combine further constraints to obtain a solution even when few images are available. Experiments on three different realistic datasets show the applicability and the potentials of the approaches.