AI Paper Atlas

Research map / Video / Video segmentation and tracking

Real-World Video Deblurring: A Benchmark Dataset and an Efficient Recurrent Neural Network

Zhihang Zhong, Ye Gao, Yinqiang Zheng, Bo Rui Zheng, Imari Sato

IJCV 2022 · Video › Video segmentation and tracking · approach: Video restoration with temporal priors · 61 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add high-level feature fusion across frames (ablation)ESTRNN (matched cost) · GOPRO (ablation, matched cost)PSNR (dB): 30.29 -> 30.46Improved
add RDB cell instead of residual block (ablation)ESTRNN (matched cost) · GOPRO (ablation, matched cost)PSNR (dB): 30.29 -> 30.51Improved
add GSA module instead of channel concatenation (ablation)ESTRNN (matched cost) · GOPRO (ablation, matched cost)PSNR (dB): 30.55 -> 30.79Improved
add fusion and RDB cell together (ablation)ESTRNN (matched cost) · GOPRO (ablation, matched cost)PSNR (dB): 30.29 -> 30.55Improved
increase RDB blocks from 3 to 15 (ablation)ESTRNN · GOPRO (RDB blocks / neighboring frames)PSNR (dB): 29.74 -> 31.27Improved
increase RDB blocks (B15C80) instead of channels (B9C90) at equal cost (ablation)ESTRNN · GOPROPSNR (dB): 31.07 -> 31.27Improved
use future frames too (F3P3 vs F0P1) (ablation)ESTRNN GSA · GOPRO (RDB blocks / neighboring frames)PSNR (dB): 30.54 -> 30.82Improved
no future frames, only past frames (F0P3) (ablation)ESTRNN GSA · GOPRO (RDB blocks / neighboring frames)PSNR (dB): 30.69 vs 30.79 (F2P2)No clear effect
increase channels C from 60 to 90 (ablation)ESTRNN · GOPROPSNR (dB): 30.12 -> 31.07Improved
train on real-world BSD instead of synthetic GOPRO, test on cross datasetESTRNN (B15C80) · Cross-validation BSD vs GOPROPSNR (dB): 26.46 on GOPRO vs 19.48 on BSDImproved
longer exposure (3ms-24ms vs 1ms-8ms)ESTRNN (B15C80) · BSDPSNR (dB): 33.36 -> 31.39Hurt
higher-frame-rate synthetic data (2000fps)Synthetic training data · BSD cross-validationqualitativeNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy