Learning classifier families for object detection and parameter estimation

The work herein described is in line with the former 3D reconstruction procedure: most of the reconstruction methods do not use image data and solely rely on the experience of the operator. Having a DRR generator would allow to perform 2D/3D registration of vertebrae models for recovering the vertebrae’s shape. However, this would translate into an optimization process, requiring the computation of hundreds of DRRs per second, thus creating a computational bottleneck when considering this task on CPUs (Russakoff et al., 2003; Ruijters et al., 2008).