YJ
Yi Jin
Author with expertise in Optical Wireless Communication Systems and Technologies
Achievements
Cited Author
Key Stats
Upvotes received:
0
Publications:
12
(33% Open Access)
Cited by:
971
h-index:
41
/
i10-index:
106
Reputation
Biology
< 1%
Chemistry
< 1%
Economics
< 1%
Show more
How is this calculated?
Publications
0

Masked Video Pretraining Advances Real-world Video Denoising

Yi Jin et al.Jan 1, 2025
Learning-based video denoisers have attained state-of-the-art (SOTA) performances on public evaluation benchmarks. Nevertheless, they typically encounter significant performance drops when applied to unseen real-world data, owing to inherent data discrepancies. To address this problem, this work delves into the model pretraining techniques and proposes masked central frame modeling (MCFM), a new video pretraining approach that significantly improves the generalization ability of the denoiser. This proposal stems from a key observation: pretraining denoiser by reconstructing intact videos from the corrupted sequences, where the central frames are masked at a suitable probability, contributes to achieving superior performance on real-world data. Building upon MCFM, we introduce a robust video denoiser, named MVDenoiser, which is firstly pretrained on massive available ordinary videos for general video modeling, and then finetuned on costful real-world noisy/clean video pairs for noisy-to-clean mapping. Additionally, beyond the denoising model, we further establish a new paired real-world noisy video dataset (RNVD) to facilitate cross-dataset evaluation of generalization ability. Extensive experiments conducted across different datasets demonstrate that the proposed method achieves superior performance compared to existing methods. Code and dataset are available at https://github.com/mxxx99/MVDenoiser .
Load More