Generation of object-boundary-matched accurate depth map is important to semi-automatic 2D to 3D conversion for high quality perception of converted 3D video. In this paper, we propose an accurate depth map generation scheme using an MRF-based initial contour tracking followed by a robust graph-cut-based contour refinement. To limit the propagation of potential tracking errors, key-frames as well as current and previous video frames are considered together in composing the graph cut seeds of the proposed contour refinement. In addition, the proposed scheme checks the ambiguity of seeds in foreground and background to eliminate the seeds, which may cause confusion in graph-cut segmentation, and thereby further improves the accuracy of contour tracking. Experiment results demonstrate the proposed scheme is especially useful when there is complex background or deformation of object, indicating the power of the proposed scheme in object-boundary-matched accurate depth map generation.