A motion vector estimation circuit (101) interpolates, with sub-pixel precision, a prediction block indicated by a motion vector estimated with integer-pixel precision as well as a region around the prediction block within a search area in a reference picture, using a 2-tap pixel interpolator (605), and estimates two or more motion vectors indicating positions of prediction blocks having high correlation with a current block to be coded, within the interpolated region. A motion compensation circuit (102) interpolates, with sub-pixel precision, each of the prediction blocks indicated by the estimated motion vectors, using a 6-tap pixel interpolator (901), compares the coding costs of the interpolated prediction blocks, and determines a motion vector indicating a position of a prediction block having highest correlation with the current block, from among the estimated motion vectors. As mentioned above, the number of motion vector candidates is narrowed down by sub-pixel precision motion prediction using a reference image created by a 2-tap filter, and then motion compensation is performed on these motion vector candidates using a 6-tap filter. Therefore, it becomes possible to achieve moving picture compression performance as high as sub-pixel precision motion prediction using a reference image created by a 6-tap filter.