TensoNRV: Tensorial Neural Representation for Videos
Yuanjie Cao, Na Li, Yun Zhang · 2025
Implicit neural representations have demonstrated strong performance in various visual tasks, such as video compression and denoising. However, existing implicit neural representation methods fail to deliver rate-quality performance comparable to video compression techniques. In this paper, we propose a novel video compression model, TensoNRV, based on the tensorial radiance fields of videos. TensoNRV performs tensor decomposition on the entire video, taking into account the spatiotemporal information of the video, thereby enhancing its representation capability. It demonstrates superior performance in tasks such as interpolation and image restoration. The use of trainable masks further boosts the compression efficiency of TensoNRV. The proposed method has been evaluated for video compression on the Bunny and UVG datasets, demonstrating significant improvements over NeRV and HNeRV. Specifically, on the UVG dataset, it achieves a $\mathbf{3 1. 6 1 \%}$ reduction in total bitrate compared to HNeRV under the PSNR metric, and a $21.83 \%$ reduction in total bitrate under the MS-SSIM metric.