An image decoder including: a demultiplexer extracting motion vector information, quantized DCT coefficients, and rounding method information; a dequantizer dequantizing to DCT coefficients; an inverse DCT converter converting to DCT coefficients to obtain an error image; a synthesizer synthesizing a prediction image of a currently decoded image by performing motion compensation using the motion vector information, the rounding method information, and a reference image which is a previously decoded image; and an adder adding the prediction image to the error image to obtain a decoded image; wherein the motion compensation is performed with half-pixel accuracy and uses bilinear interpolation to calculate intensity values of chrominance or luminance at points where no pixels actually exist in the reference image, the bilinear interpolation being performable using positive and negative rounding methods and is performed using a rounding method specified by the rounding method information specifying one of a plurality of values.