In Inverse Synthetic Aperture Radar (ISAR) imaging, phase error caused by three-dimensional rotation is relevant to the position of the scatterer, so it is difficult to use traditional phase autofocus algorithms to complete phase compensation, in which the uniform phase correction function is applied. In order to resolve this problem, a three-dimensional rotation detection method is put forward to complete the selection of imaging times, in which the motion state of target is two-dimensional rotation. In this method, the phases of two arbitrary scatterers are estimated by adaptive Gaussian chirplet decomposition (AGCD) fast algorithm, and the target rotation state is obtained by measuring the non-linearity of the two phases, which means those intervals that the target just has two-dimensional rotation are selected to obtain the ISAR images. Numerical simulation results prove its good efficiency.