GNeRV: A Global Embedding Neural Representation For Videos
Keywords: INR, video compression, progressive pipeline
Abstract: In recent years, Implicit Neural Representation (INR) has garnered considerable attention for its effectiveness in compressing various visual information while delivering significant advantages in decoding speed. Video compression work with INR use time index as input and corresponding frame in RGB format as output. However, related work suffers from poor representation performance due to insufficient information in the embedding structure. In this paper, we introduce a global embedding structure, whose parameters are generated by random initialization and back propagation without any other constraint, and this embedding is shared by all frames. Furthermore, we propose a progressive training pipeline wherein large models are built upon the reuse and expansion of small models. Our Global embedding Neural Representation for Videos (GNeRV) achieves SOTA results on multiple datasets. Taking UVG dataset as an example, GNeRV model outperforms the previously leading model HiNeRV by 1.5-2 dB at the same bitrate. And our progressive pipeline can effectively reduce the computational complexity of multi-bitrate encoding and save the storage space of multi-bitrate compressed files.
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3417
Loading