Learning to compress videos without computing motion

作者：

Highlights：

• MOVI-Codec avoids the overhead of motion estimation/compensation based on search.

• Our LTSM-UNet efficiently captures spatio-temporal information for reconstruction.

• MOVI-Codec is collectively jointly optimized using a single loss function.

• Our codec outperforms the veryfast setting of H.264 and H.265 against MS-SSIM.

摘要

•MOVI-Codec avoids the overhead of motion estimation/compensation based on search.•Our LTSM-UNet efficiently captures spatio-temporal information for reconstruction.•MOVI-Codec is collectively jointly optimized using a single loss function.•Our codec outperforms the veryfast setting of H.264 and H.265 against MS-SSIM.

论文关键词：Video compression,Deep learning,Motion

论文评审过程：Received 5 July 2021, Revised 29 December 2021, Accepted 3 January 2022, Available online 6 January 2022, Version of Record 31 January 2022.

论文官网地址：https://doi.org/10.1016/j.image.2022.116633