for 3D objects it is extremely expensive to exhaustively consider all possible candidates. Under a uniform sampling scheme over the whole parameter space, the number of tested candidates exponentially increases with the dimensionality of the parameter space. In our fetal face detection system, we used a marginal space learning framework proposed by Zheng, et al [16], which basically learns the location parameters in a sequential way on projected sampling sub-spaces. A classifier is trained for each projected space using the probabilistic boosting-tree (PBT) [11].