Neural Residual Flow Fields for Efficient Video Representations

TL;DR

This paper introduces a neural field architecture that leverages motion information and multiple reference frames to efficiently represent and compress videos, significantly reducing parameters while improving performance.

Contribution

It proposes a novel neural residual flow field method that exploits motion redundancy and multiple references for more parameter-efficient video representation and compression.

Findings

Outperforms baseline methods significantly

Uses fewer parameters due to motion-based redundancy removal

Effective with multiple reference frames

Abstract

Neural fields have emerged as a powerful paradigm for representing various signals, including videos. However, research on improving the parameter efficiency of neural fields is still in its early stages. Even though neural fields that map coordinates to colors can be used to encode video signals, this scheme does not exploit the spatial and temporal redundancy of video signals. Inspired by standard video compression algorithms, we propose a neural field architecture for representing and compressing videos that deliberately removes data redundancy through the use of motion information across video frames. Maintaining motion information, which is typically smoother and less complex than color signals, requires a far fewer number of parameters. Furthermore, reusing color values through motion information further improves the network parameter efficiency. In addition, we suggest using more than one reference frame for video frame reconstruction and separate networks, one for optical flows and the other for residuals. Experimental results have shown that the proposed method outperforms the baseline methods by a significant margin. The code is available in https://github.com/daniel03c1/eff_video_representation