Decoder-side motion vector refinement (DMVR) process method and apparatus

An inter prediction apparatus, comprising:

• a non-transitory memory having processor-executable instructions stored thereon;
• and
• one or more processors, coupled to the memory, configured to execute the processor-executable instructions to cause the inter prediction apparatus to
• obtain sub-pixel accurate merge motion vectors of a current block;
• obtain motion compensated interpolated samples based on the sub-pixel accurate merge motion vectors from a bilinear motion compensated interpolation;
• compute a sum of absolute differences (SAD) between two motion compensated reference patches using a subset of the motion compensated interpolated samples, wherein the number of samples in the subset of the motion compensated interpolated samples is less than the number of samples in the motion compensated interpolated samples;
• determine whether the SAD is less than a coding unit (CU) size-dependent threshold value;
• when the SAD is less than the CU size-dependent threshold value:
• skip remaining decoder-side motion vector refinement (DMVR) process steps;
• and
• perform final motion compensation based on the sub-pixel accurate merge motion vectors of the current block;
• and
• when the SAD is greater than or equal to the CU size-dependent threshold value:
• perform the remaining DMVR process steps;
• and
• perform the final motion compensation based on the sub-pixel accurate merge motion vectors of the current block and motion vector corrections;
• wherein the remaining DMVR process steps comprises:
• perform integer distance refinement initially at integer distance from the sub-pixel accurate merge motion vectors, to obtain cost function values at integer distance positions, wherein a position with a lowest cost function value corresponds to a final integer distance;
• obtain sub-pixel distance based on the cost function values at the integer distance positions and the cost function value at the position with the lowest cost function value;
• obtain the motion vector corrections based on the final integer distance and the sub-pixel distance.

A non-transitory computer-readable medium storing processor-executable instructions for one or more processors, when executed by the one or more processors, causing the one or more processors to perform steps of

• obtaining sub-pixel accurate merge motion vectors of a current block;
• obtaining motion compensated interpolated samples based on the sub-pixel accurate merge motion vectors from a bilinear motion compensated interpolation;
• computing a sum of absolute differences (SAD) between the two motion compensated reference patches using a subset of the motion compensated interpolated samples, wherein the number of samples in the subset of the motion compensated interpolated samples is less than the number of samples in the motion compensated interpolated samples;
• determining whether the SAD is less than a coding unit (CU) size-dependent threshold value;
• when the SAD is less than the CU size-dependent threshold value:
• skipping remaining decoder-side motion vector refinement (DMVR) process steps;
• and
• performing final motion compensation based on the sub-pixel accurate merge motion vectors of the current block;
• and
• when the SAD is greater than or equal to the CU size-dependent threshold value:
• performing the remaining DMVR process steps;
• and
• performing the final motion compensation based on the sub-pixel accurate merge motion vectors of the current block and motion vector corrections;
• wherein the DMVR process steps comprises:
• performing integer distance refinement initially at integer distance from the sub-pixel accurate merge motion vectors, to obtain cost function values at integer distance positions, wherein a position with a lowest cost function value corresponds to a final integer distance;
• obtaining sub-pixel distance based on the cost function values at the integer distance positions and the cost function value at the position with the lowest cost function value;
• obtaining the motion vector corrections based on the final integer distance and the sub-pixel distance.