Segmenting neural video representation for low startup delay

Abstract

Implicit Neural Representation(INR) is the method that represents data using a neural network. Its strong modeling power and advantage of continuous representation make INR can successfully store information of various data types. However, depending on the nature of information representation, the larger the information needed to be stored, the larger the size of the INR model to maintain the same representation performance, which results in a large start delay in the aspect of network delivery. We suggest a new INR model segmenting method that can successfully reduce the startup delay that occurs transfer the INR model through the network while minimizing the negative effects on the data reconstruction quality. The experiment shows that our segmenting methods can reduce the startup delay x2 ~ x4 with minimized PSNR drop (<= 0.5dB)