A new and robust mapping approach is proposed entitled mapping forests (MFs) for computer vision applications based on regression transformations. Mapping forests relies on learning nonlinear mappings deduced from pairs of source and target training data, and improves the performance of mappings by enabling nonlinear transformations using forests. In contrast to previous approaches, it provides automatically selected mappings, which are naturally nonlinear. MF can provide accurate nonlinear transformations to compensate the gap of linear mappings or can generalize the nonlinear mappings with constraint kernels. In our experiments, we demonstrate that the proposed MF approach is not only on a par or better than linear mapping approaches and the state-of-the-art, but also is very time efficient, which makes it an attractive choice for real-time applications. We evaluated the efficiency and performance of the MF approach using the BU3DFE and Multi-PIE datasets for multi-view facial expression recognition application, and Set5, Set14 and SuperTex136 datasets for single image super-resolution application.