Neural Residual Flow Fields for Efficient Video Representations

Neural fields have emerged as a powerful paradigm for representing various signals, including videos. However, research on improving the parameter efficiency of neural is still in its early stages. Even though that map coordinates to colors can be used encode video this scheme does not exploit spatial and temporal redundancy signals. Inspired by standard compression algorithms, we propose field architecture compressing videos deliberately removes data through use motion information across frames. Maintaining information, which typically smoother less complex than color requires far fewer number parameters. Furthermore, reusing values further improves network efficiency. In addition, suggest using more one reference frame reconstruction separate networks, optical flows other residuals. Experimental results shown proposed method outperforms baseline methods significant margin.