Research map / Label-efficient and robust learning / Domain adaptation and generalization
ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance
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 ABC-Former, a bimodal transformer that uses two histogram-based auxiliary PDFformers (sRGB and CIELab) to guide an sRGB target transformer via Interactive Channel Attention for white balance correction.
Problem
Correcting color casts and inaccurate white balance in sRGB images produced by camera ISPs, where existing methods ignore global color trends and rely on local sRGB information.
Approach
Converts images into sRGB and CIELab color-probability histograms processed by auxiliary 1D transformers (PDFformers); their corrected global color knowledge is transferred into a U-shaped sRGB transformer through re-parameterized channel attention (ICA) weights in each block.
Key results
- On Rendered WB Set1-Test, ABC-Former reports mean MSE 20.47 and mean MAE 1.99 degrees, versus 78.89 and 2.67 for the next best WBFlow.
- On Rendered Cube+, ABC-Former reaches mean MSE 60.60, MAE 2.99 degrees and delta E 2000 3.95, the lowest among compared methods.
- Ablation on Cube+ shows the full two-auxiliary design (MSE 60.60) beats the sRGB-only target model (MSE 76.56), and L2 histogram loss beats KL divergence and Wasserstein.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ABC-Former instead of KNN | White balance correction · Set1-Test | MSE (mean): 77.49 -> 20.47 | Helped |
| Use ABC-Former instead of WBFlow | White balance correction · Set2 | MSE (mean): 117.60 -> 104.31 | Helped |
| Use ABC-Former instead of SWBNet | White balance correction · Rendered Cube+ | MSE (mean): 74.35 -> 60.60 | Helped |
| Add the CIELab histogram auxiliary model (ablation) | sRGBformer · Rendered Cube+ | MSE (mean): 76.56 -> 73.35 | Helped |
| Add the CIELab histogram auxiliary model (ablation) | sRGBformer · Rendered Cube+ | MAE (mean): 3.35° -> 3.26° | Helped |
| Add the CIELab histogram auxiliary model (ablation) | sRGBformer · Rendered Cube+ | ∆E 2000 (mean): 4.31 -> 4.20 | Helped |
| Add the sRGB histogram auxiliary model (ablation) | sRGBformer · Rendered Cube+ | MSE (mean): 76.56 -> 68.65 | Helped |
| Add the sRGB histogram auxiliary model (ablation) | sRGBformer · Rendered Cube+ | MAE (mean): 3.35° -> 3.12° | Helped |
| Add the sRGB histogram auxiliary model (ablation) | sRGBformer · Rendered Cube+ | ∆E 2000 (mean): 4.31 -> 4.08 | Helped |
| Use one auxiliary model for both modalities rather than separate sRGB and CIELab models (ablation) | ABC-Former · Rendered Cube+ | MSE (mean): 60.60 -> 72.38 | Hurt |
| Use one auxiliary model for both modalities rather than separate sRGB and CIELab models (ablation) | ABC-Former · Rendered Cube+ | MAE (mean): 2.99° -> 3.38° | Hurt |
| Use one auxiliary model for both modalities rather than separate sRGB and CIELab models (ablation) | ABC-Former · Rendered Cube+ | ∆E 2000 (mean): 3.95 -> 4.38 | Hurt |
Surprising findings
- A single auxiliary model jointly learning sRGB and CIELab histograms performed worse than using two separate auxiliary models.
Limitations
- Compared SWBNet was retrained by the authors on Set1-Test and Set2 because original results were unavailable.
- Histograms are per-channel and do not capture pixel-level color relationships, as the authors note.
Experimental setup
Datasets: Rendered WB dataset Set1 (training folds)
Benchmarks: Rendered WB Set1-Test Rendered WB Set2 Rendered Cube+
Metrics: MSE MAE delta E 2000 Bhattacharyya coefficient
Baselines: KNN Deep-WB Mixed-WB WBFlow SWBNet
Builds on: M2PT KNN Deep-WB WBFlow SWBNet
Closest papers by meaning
- HVDualformer: Histogram-Vision Dual Transformer for White Balance (AAAI 2025 · similarity 0.86)
- White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation (CVPR 2026 · similarity 0.85)
- Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing (CVPR 2024 · similarity 0.82)
- EnColor: Improving Visual Accessibility with a Deep Encoder-Decoder Image Corrector for Color Vision Deficient Individuals (AAAI 2024 · similarity 0.82)
- Color Shift Estimation-and-Correction for Image Enhancement (CVPR 2024 · similarity 0.82)
- Efficient RAW Image Deblurring with Adaptive Frequency Modulation (NeurIPS 2025 · similarity 0.82)
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.