Research map / Architectures and efficiency / RNNs and normalization
Integral Fast Fourier Color Constancy
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.
Extends the FFCC color-constancy method to multi-illuminant scenes using an integral UV histogram, parallel Fourier-domain prediction over overlapping windows, and guided-filter smoothing, targeting real-time CPU auto white balance.
Problem
Auto white balance in multi-illuminant scenes: existing neural pixel-level methods are heavy and slow for real-time use, while FFCC handles only a single global illuminant and needs repeated histogram extraction per region.
Approach
Precompute an integral log-chroma UV histogram so any window's histogram is obtained with a few additions/subtractions, predict illuminants for all overlapping windows in parallel via Fourier-domain convolution and circular-mean BVM decoding, then smooth with linear interpolation and guided filtering to get a continuous illumination map.
Key results
- On the Shadow dataset (Canon 5d subset), IFFCC reaches mean angular error 2.06 vs 2.63 for Domislovic et al. and 3.14 for LSMI-U.
- On LSMI multi-illuminant images, IFFCC achieves mean 2.48 (Galaxy) with 0.012M parameters and 0.03 s CPU time for 256x256 images, versus AID at 2.03 with 6.4M parameters and >1 s.
- IFFCC runs in 5.8 ms on 64x48 preview images on a CPU, per the paper's claim of 20-100x speedup over network-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Extend FFCC with integral UV histograms, parallel prediction, and spatial smoothing | FFCC · Shadow, use all | mean angular error: 2.28 -> 2.19 | Helped |
| Extend FFCC with integral UV histograms, parallel prediction, and spatial smoothing | FFCC · Shadow, outdoor | mean angular error: 2.12 -> 1.74 | Helped |
| Extend FFCC with integral UV histograms, parallel prediction, and spatial smoothing | FFCC · Shadow, indoor | mean angular error: 8.78 -> 3.04 | Helped |
| Extend FFCC with integral UV histograms, parallel prediction, and spatial smoothing | FFCC · LSMI, Multi | mean angular error: 2.03 -> 2.48; 2.36 -> 2.30; 2.16 -> 2.48 | Mixed |
| Extend FFCC with integral UV histograms, parallel prediction, and spatial smoothing | FFCC · LSMI, Mixed | mean angular error: 1.63 -> 1.98; 2.16 -> 2.11; 1.68 -> 1.93 | Mixed |
| Use window 32 and overlap 16 instead of window 128 and overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.65 | Hurt |
| Use window 64 and no overlap instead of window 128 and overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.35 | Hurt |
| Use dominant-light GT and window 64 with overlap 32 instead of blended GT and window 128 with overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.53 | Hurt |
| Use window 64 and overlap 32 instead of window 128 and overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.37 | Hurt |
| Use dominant-light GT and no overlap instead of blended GT and overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.46 | Hurt |
| Use no overlap instead of overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.16 | Hurt |
| Use dominant-light GT and overlap 32 instead of blended GT and overlap 64 (ablation) | IFFCC · Shadow, Canon5d | mean angular error: 2.06 -> 2.50 | Hurt |
Surprising findings
- Integral-histogram IFFCC on the Shadow dataset outperforms several network-based methods, and the paper reports it is state-of-the-art among the compared methods there.
Limitations
- Larger windows blur boundaries in multi-illuminant regions, while smaller windows lose accuracy.
- Interpolation may blur local details at strong light edges, which guided filtering only partly mitigates.
Experimental setup
Datasets: LSMI dataset Shadow dataset
Benchmarks: Shadow dataset LSMI dataset
Metrics: mean angular error median angular error best 25% mean worst 25% mean parameter count CPU inference time
Baselines: FFCC LSMI-U AID Domislovic et al. Bianco et al. Pix2Pix Gijsenij CRF (White-Patch) Patch-based HypNet/SelNet
Builds on: Fast Fourier Color Constancy (FFCC) Convolutional Color Constancy (CCC) LSMI guided filtering
Closest papers by meaning
- Fast Fourier Color Constancy (CVPR 2017 · similarity 0.90)
- A Multi-Hypothesis Approach to Color Constancy (CVPR 2020 · similarity 0.84)
- Attentive Illumination Decomposition Model for Multi-Illuminant White Balancing (CVPR 2024 · similarity 0.83)
- GCC: Generative Color Constancy via Diffusing a Color Checker (CVPR 2025 · similarity 0.81)
- White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation (CVPR 2026 · similarity 0.81)
- ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention for White Balance (CVPR 2025 · similarity 0.81)
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.