Controllable-quality neural video representation

With the rapid growth of video data, efficient and high-quality video compression has become increasingly important in computer vision. Inspired by implicit neural representations, neural representation-based video compression methods have shown promise in compressing video data by fitting frames with neural networks. Building on this idea, we propose MR-NeRV, a novel video representation framework that models inter-frame motion and residuals rather than raw pixel values. This approach reduces the complexity of the fitting process and enhances the reconstruction of high-frequency details. Based on the MR-NeRV framework, we propose the first controllable-quality video reconstruction algorithm for implicit neural representations. This algorithm enable videos to be reconstructed at varying quality levels using a single model by adjusting the number and distribution of compressed frames. To further reduce model size, we propose the Compressed Neural Layer, which substantially reduces the network parameter count while preserving visual quality. Experimental results show that MR-NeRV achieves a 7.4% performance improvement over the baseline, delivering performance on par with advanced algorithms. It outperforms them in the reconstruction of high-frequency details. Moreover, the proposed controllable-quality reconstruction approach achieves an average 63% reduction in training time compared to state-of-the-art algorithms. To support more comprehensive evaluation, we curate and annotate a new collection of high-quality animation videos, providing a valuable addition to current datasets.