The motion compensation prediction using a virtual sample is used for detection of a motion vector in each area and the virtual sample accuracy is locally decided by relating it to the area size of each motion vector detection unit. For example, in the motion vector detection unit area of sizes 8x4, 4x8, 4x4 smaller than 8x8 MC, a virtual sample of half pixel accuracy is used, and in the motion vector detection unit area of 8x8 MC or above, a virtual sample of quarter pixel accuracy is used. Thus, it is possible to reduce the memory band width required for generating a virtual sample for each of the modes 8x4, 4x8, 4x4 MC and simplify the device.