US 12,149,726 B2
Motion refinement using a deep neural network
Franck Galpin, Cesson-Sevigne (FR); Philippe Bordes, Cesson-Sevigne (FR); Philippe Guillotel, Cesson-Sevigne (FR); and Xuan Hien Pham, Fontenay-aux-Roses (FR)
Assigned to InterDigital Madison Patent Holdings, SAS, Paris (FR)
Appl. No. 17/921,293
Filed by InterDigital Madison Patent Holdings, SAS, Paris (FR)
PCT Filed May 18, 2021, PCT No. PCT/EP2021/063056
§ 371(c)(1), (2) Date Oct. 25, 2022,
PCT Pub. No. WO2021/239500, PCT Pub. Date Dec. 2, 2021.
Claims priority of application No. 20305565 (EP), filed on May 29, 2020.
Prior Publication US 2023/0171421 A1, Jun. 1, 2023
Int. Cl. H04N 19/51 (2014.01)
CPC H04N 19/51 (2014.11) 20 Claims
OG exemplary drawing
 
1. A method for video decoding, comprising:
obtaining a first motion field and a second motion field for a block of a picture, wherein said first motion field corresponds to a first reference picture and said second motion field corresponds to a second reference picture;
obtaining a first motion-compensated prediction block for said block based on said first motion field for said block, and obtaining a second motion-compensated prediction block for said block based on said second motion field for said block;
obtaining a third motion field representative of motion between said first and second motion-compensated prediction blocks, using a deep neural network, wherein said third motion field is a pixel-based motion field;
refining said first and second motion fields, based on said third motion field;
obtaining a prediction block for said block, based on said refined first and second motion fields; and
decoding said block based on said prediction block.