Decompose More and Aggregate Better: Two Closer Looks at Frequency Representation Learning for Human Motion Prediction | Read Paper on Bytez