Research map / Image generation / Image restoration
Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoTF, a real-time exposure correction framework that combines image-adaptive 3D LUT global transformation with low-resolution pixel-wise transformation via a relation-aware modulation module, plus an adaptive sampling strategy that learns non-uniform downsampling grids for the LUT input.
Problem
Exposure correction methods using dense pixel-wise transformations are computationally expensive at high resolution, while 3D LUT methods are efficient but only perform global transformations and downsample uniformly, losing color information.
Approach
Collaboratively fuse a learnable 3D LUT global transformation (fed by a content-adaptive sampling grid that densely samples colorful regions) with a low-resolution encoder-decoder pixel transformation, using a cross-resolution relation-aware modulation module to inject local context into the global result.
Key results
- On MSEC, CoTF reaches 23.44 dB PSNR, 0.8728 SSIM and 0.1232 LPIPS, the best among compared methods.
- On LCDP, CoTF improves PSNR by 0.65 dB over the second-best LCDPNet (23.89 vs 23.24 dB) while using 1.81 GFLOPs versus 27.12 GFLOPs at 1024x1024.
- Adaptive sampling adds only about 0.02G FLOPs at 256x256 and improves 3D LUT performance across LUT sizes and sampling resolutions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoTF instead of FECNet+ERL | Exposure correction · MSEC average | PSNR: 23.15 -> 23.44 | Helped |
| Use CoTF instead of FECNet+ERL | Exposure correction · SICE average | PSNR: 21.22 -> 21.51 | Helped |
| Use CoTF instead of LCDPNet | Exposure correction · LCDP | PSNR: 23.24 -> 23.89 | Helped |
| Use only low-resolution pixel-wise transformation (ablation) | CoTF · LCDP | PSNR: 20.08 -> 23.89 | Hurt |
| Use pixel-wise transformation at high resolution (ablation) | CoTF · LCDP | PSNR: 22.34 -> 23.89 | Hurt |
| Add adaptive sampling (ablation) | 3D LUT-based global transformation · LCDP | PSNR: 23.07 -> 23.89 | Helped |
| Replace concatenation after upsampling with RAM (ablation) | CoTF · LCDP | PSNR: 23.43 -> 23.89 | Helped |
| Replace channel attention after upsampling with RAM (ablation) | CoTF · LCDP | PSNR: 23.62 -> 23.89 | Helped |
| Replace RAM* after upsampling with RAM (ablation) | CoTF · LCDP | PSNR: 23.79 -> 23.89 | Helped |
| Use adaptive sampling during training only (ablation) | 3D LUT-based global transformation · LCDP | PSNR: 23.07 -> 23.33 | Helped |
| Enable adaptive sampling at test time after training with it (ablation) | 3D LUT-based global transformation · LCDP | PSNR: 23.33 -> 23.35 | Helped |
Surprising findings
- Adaptive sampling used only as a training-time data augmentation (deactivated at test time) still improves PSNR from 23.07 to 23.33 dB.
Limitations
- Global 3D LUT ignores local pixel context, which is addressed only by the added low-resolution pixel branch.
- Adaptive sampling is only applied to the global branch; pixel branch still uses bilinear downsampling.
- UHD results are slower than pure 3D LUT (0.0431 s vs 0.0023 s) and PSNR is only 19.09 dB.
Experimental setup
Datasets: MSEC SICE LCDP
Benchmarks: MSEC SICE LCDP
Metrics: PSNR SSIM LPIPS Params FLOPs runtime
Baselines: MSEC SID DRBN ENC-SID ENC-DRBN FECNet LCDPNet CLIP-LIT FECNet+ERL Zero-DCE Zero-DCE++ RUAS
Builds on: DeepLUT (image-adaptive 3D LUTs) AdaInt ENC MSEC Zero-DCE
Closest papers by meaning
- Color Shift Estimation-and-Correction for Image Enhancement (CVPR 2024 · similarity 0.86)
- NILUT: Conditional Neural Implicit 3D Lookup Tables for Image Enhancement (AAAI 2024 · similarity 0.85)
- Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment (CVPR 2025 · similarity 0.85)
- High-Resolution Image Harmonization with Adaptive-Interval Color Transformation (NeurIPS 2024 · similarity 0.84)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.84)
- GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation (NeurIPS 2026 · similarity 0.84)
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.