Recently, marginal space learning (MSL) was proposed as a generic approach for automatic detection of 3D anatomical structures in many medical imaging modalities [1]. To accurately localize a 3D object, we need to estimate nine pose parameters (three for position, three for orientation, and three for anisotropic scaling). Instead of exhaustively searching the original nine-dimensional pose parameter space, only low-dimensional marginal spaces are searched in MSL to improve the detection speed. In this paper, we apply MSL to 2D object detection and perform a thorough comparison between MSL and the alternative full space learning (FSL) approach. Experiments on left ventricle detection in 2D MRI images show MSL outperforms FSL in both speed and accuracy. In addition, we propose two novel techniques, constrained MSL and nonrigid MSL, to further improve the efficiency and accuracy. In many real applications, a strong correlation may exist among pose parameters in the same marginal spaces. For example, a large object may have large scaling values along all directions. Constrained MSL exploits this correlation for further speed-up. The original MSL only estimates the rigid transformation of an object in the image, therefore cannot accurately localize a nonrigid object under a large deformation. The proposed nonrigid MSL directly estimates the nonrigid deformation parameters to improve the localization accuracy. The comparison experiments on liver detection in 226 abdominal CT volumes demonstrate the effectiveness of the proposed methods. Our system takes less than a second to accurately detect the liver in a volume.