Abstract: For a bi-prediction block, the initial motion field can be refined using a DNN. In one implementation, the initial motion field is integer rounded to obtain initial prediction blocks. Based on the initial prediction, the DNN can generate motion refinement information, which is scaled and added to the sub-pel residual motion from the initial motion field to generate a refined motion field. The scaling factor can take a default value, or be based on the motion asymmetry. While the initial motion field is usually block based on sub-block based, the refined motion field is pixel based or sub-block based and can be at an arbitrary accuracy. The same refinement process is performed at both the encoder and decoder, and thus the motion refinement information need not to be signaled. Whether the refinement is enabled can be determined based on the initial motion, the block activities and the block size.

Abstract: At least a method and an apparatus are presented for efficiently encoding or decoding video. For example, a plurality of frames is provided to a motion estimator to produce an output comprising estimated motion information. The estimated motion information is provided to an auto-encoder or an auto-decoder to produce an output comprising reconstructed motion field. The reconstructed motion field and one or more decoded frames of the plurality of frames are provided to a deep neural network to produce an output comprising refined bi-directional motion field. The video is encoded or decoded based on the refined bi-directional motion field.