STFNet: Scalable Blind RAW Video Demosaicing and Denoising via Stable Time-Travel Fusion
Year: 2026
Type: conference-paper
Abstract: Blind RAW video demosaicing and denoising remains challenging because RAW imaging is camera dependent, and collecting large-scale paired RAW video datasets is prohibitively expensive. In this paper, we present a scalable synthetic-only framework for blind RAW video demosaicing and denoising. Instead of relying on paired RAW video supervision, our framework learns from physics-based synthetic RAW videos whose inverse imaging process and degradations preserve both camera realism and temporal coherence. Specifically, we develop a RAW video synthesis pipeline that converts clean sRGB videos into degraded Bayer RAW sequences using real camera parameters, along with temporally shared sensor noise. On the model side, we propose STFNet, a flow-free causal demosaicing and denoising framework that combines Stable Temporal Propagation and Historical Temporal Aggregation to improve temporal reliability under online inference. Extensive experiments suggest that, under current benchmarks, synthetic-only training transfers competitively to real RAW videos while achieving a favorable balance between restoration quality and computational efficiency. In addition, we introduce Ultra-RAW, a synthetic benchmark with exact paired ground truth for full-reference evaluation, while using real benchmarks as transfer testbeds for cross-camera assessment. Code is available at https://github.com/nsqg/STFNet.
Authors: Qingguo Liu, Qianjun Huang, Kai Zhang, Jian Yang
Institution: Nanjing University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225416638