Research map / 3D vision / Depth and stereo
Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing
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 AID, a slot-attention network that predicts per-illuminant chromaticities and weight maps for multi-illuminant white balance, trained with a centroid-matching loss.
Problem
Multi-illuminant white balance: existing encoder-decoder methods predict pixel-level illumination directly, ignoring the number and chromaticity of light sources, which yields inconsistent illumination maps.
Approach
Use slot attention so each slot represents one illuminant, with its attention map as the pixel weight map and a chromaticity MLP output; the mixed illumination map is their linear combination per the Lambertian model. A centroid-matching loss (Hungarian-style matching to K-means color centroids) activates the right slots.
Key results
- On LSMI all-in-one setting, AID reaches mean MAE 2.04 vs 2.31 for LSMI-U; with multi-domain learning (AID + MDL) 1.93.
- On MIIW mixed-illuminant test set, AID reduces mean MAE from 4.28 (LSMI-U) to 2.46.
- On NUS-8 three-fold cross-validation, AID achieves mean angular error 1.57 vs 1.84 for CLCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AID instead of LSMI-U | LSMI all-in-one cross-camera · LSMI all-in-one cross-camera | mean angular error (MAE): 2.31 -> 2.04 | Helped |
| Add camera-specific slot initialization (MDL) | AID · LSMI all-in-one cross-camera | mean angular error (MAE): 2.04 -> 1.93 | Helped |
| Use AID instead of LSMI-U | MIIW · MIIW single, multi, and mixed illumination | mean angular error (MAE): 4.15 -> 1.07; 4.34 -> 3.14; 4.28 -> 2.46 | Helped |
| Use AID instead of CLCC | NUS-8 · NUS-8 | Mean: 1.84 -> 1.57 | Helped |
| Add centroid-matching loss (ablation) | AID · LSMI Galaxy | illuminant number accuracy (# acc.): 0.288 -> 0.800 | Helped |
| Add centroid-matching loss (ablation) | AID · LSMI Galaxy | mean angular error (MAE): 1.58 -> 1.66 | Hurt |
| Use 5 rather than 7 slots (ablation) | AID · LSMI Galaxy | mean angular error (MAE): 1.66 -> 1.82 | Hurt |
| Use 9 rather than 7 slots (ablation) | AID · LSMI Galaxy | mean angular error (MAE): 1.66 -> 1.84 | Hurt |
| Use 2 rather than 3 slot-attention iterations (ablation) | AID · LSMI Galaxy | mean angular error (MAE): 1.66 -> 1.85 | Hurt |
| Use 4 rather than 3 slot-attention iterations (ablation) | AID · LSMI Galaxy | mean angular error (MAE): 1.66 -> 1.92 | Hurt |
Limitations
- Requires presetting the number of slots/clusters K.
- Number of illuminants cannot be determined dynamically from the input.
Experimental setup
Datasets: LSMI Multi-Illumination In the Wild (MIIW) NUS-8
Benchmarks: LSMI MIIW NUS-8
Metrics: mean angular error median angular error tri-mean best 25% worst 25% geometric mean illuminant count accuracy illuminant angular error
Baselines: LSMI-U LSMI-H AWB Patch CNN AngularGan TransCC CCC AlexNet-FC4 FFCC CLCC
Builds on: slot attention FC4 LSMI
Closest papers by meaning
- Two Illuminant Estimation and User Correction Preference (CVPR 2016 · similarity 0.89)
- White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation (CVPR 2026 · similarity 0.86)
- GCC: Generative Color Constancy via Diffusing a Color Checker (CVPR 2025 · similarity 0.85)
- Integral Fast Fourier Color Constancy (CVPR 2025 · similarity 0.83)
- Do Image Editing Models Understand Lighting? (NeurIPS 2026 · similarity 0.83)
- Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment (CVPR 2025 · similarity 0.83)
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.