Research map / Image generation / Super-resolution
CrossNet: An End-to-end Reference-based Super Resolution Network using Cross-scale Warping
Haitian Zheng, Mengqi Ji, Haoqian Wang, Yebin Liu, Lu Fang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace multi-scale feature-domain warping with single-scale image warping before the encoder (ablation) | CrossNet · Flower x8 (1,1) | PSNR: 40.31 vs 40.14 dB (CrossNet-iw) | Improved |
| replace feature warping with image warping (ablation) | CrossNet · Flower x8 (7,7) | PSNR: 34.37 vs 34.49 dB (CrossNet-iw) | Hurt |
| replace feature warping with image warping (ablation) | CrossNet · LFVideo x8 (1,1) | PSNR: 41.26 vs 41.12 dB (CrossNet-iw) | Improved |
| replace feature warping with image warping (ablation) | CrossNet · LFVideo x8 (7,7) | PSNR: 36.48 vs 36.32 dB (CrossNet-iw) | Improved |
| pretrain flow estimator with image warping task before joint training (CrossNet-iw-p) (ablation) | CrossNet-iw · Flower x8 (1,1) | PSNR: 40.14 -> 40.01 | No clear effect |
| feature-domain warping converges faster than image-domain warping (ablation) | CrossNet · training set PSNR x8 Flower | PSNR: +0.20dB at end of training (0.27dB on LFVideo) | Improved |
| add upscaling layers to FlowNetS (FlowNetS+, 2% more parameters) (ablation) | FlowNetS flow estimator · Flower x8 (7,7) | PSNR: 31.23 -> 31.66 | Improved |
| add upscaling layers to FlowNetS (ablation) | FlowNetS flow estimator · LFVideo x8 (1,1) | PSNR: 39.39 -> 39.37 | No clear effect |
| use reference image with CrossNet instead of SISR methods | RefSR · Flower/LFVideo x4 and x8 | PSNR: 2dB-4dB gain in general | Improved |
| increase parallax between reference and LR image | RefSR methods · Flower x8 (1,1) vs (7,7) | PSNR: 40.31 -> 34.37 (CrossNet) | Hurt |
| add parallax augmentation ([-15,15] pixel reference offsets) and retrain | CrossNet · Stanford light field x8 | PSNR: 38.44 vs 35.96 (SS-Net) | Improved |
Closest papers by meaning
- CrossNet++: Cross-Scale Large-Parallax Warping for Reference-Based Super-Resolution (TPAMI 2020 · similarity 0.92)
- Image Super-Resolution by Neural Texture Transfer (CVPR 2019 · similarity 0.89)
- Robust Reference-Based Super-Resolution With Similarity-Aware Deformable Convolution (CVPR 2020 · similarity 0.89)
- Reference-based Image and Video Super-Resolution via $C^{2}$-Matching (TPAMI 2022 · similarity 0.89)
- LMR: A Large-Scale Multi-Reference Dataset for Reference-Based Super-Resolution (ICCV 2023 · similarity 0.88)
- Robust Reference-Based Super-Resolution via C2-Matching (CVPR 2021 · similarity 0.88)
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.
