SAFE: Scene-Aware Feature Modulation for Color Constancy with Learned Color Sapce in Pure-Color Scenes

Abstract

Overview of the proposed color constancy framework. Training proceeds in four phases: (1) pre-train the SAFE backbone $\phi$ alongside a coarse global $w$-module under the angular loss; (2) freeze $\phi^{(1)}$ and search for a scene-dependent color axis that minimizes the teacher’s angular error; (3) supervise the LCS predictor $\psi$ on the oracle in logit space; (4) jointly fine-tune $\phi$ and $\psi$ end-to-end. At inference, the LCS predictor $\psi$ outputs a per-image color axis $\mathbf{w}(\mathbf{I})$ that re-parameterizes the chromaticity normalization for calculating the four-token illumination feature $\mathbf{f}_w$; the SAFE backbone then modulates $\mathbf{f}_w$ via a scene descriptor to predict the illuminant estimate $\hat{\mathbf{y}}$.

Contributions

A four-token illumination feature designthat combines physics-derived cues (specular log-chromaticity, dichromatic principal direction) with chromaticity statistics from original and edge images into a structured 24D representation with clear semantic roles.
The Scene-Aware Feature Modulation(SAFE) network that adaptively reweights illumination tokens based on scene characteristics, enabling robust and stable behavior across diverse imaging conditions.
The Learned Color Space(LCS): a scene-dependent chromaticity normalization predicted by a lightweight model from the image and scene complexity features, designed to counter chromaticity collapse on pure-color scenes.
Experiment results on the public benchmark demonstrate that SAFE consistently improves performance in pure-color scenes. Compared to the best-performing baseline in each metric, it reduces the mean angular error by 10%, the best-25% error by 20%, and the worst-25% error by 5.8%.

Results

Table 1: Evaluation results on the PolyUv2 pure color dataset. We report the mean angular error (in degrees) based on 3-fold cross-validation. SAFE outperforms existing methods, yielding the lowest mean, median, tri-mean, best-25%, and worst-25% angular errors, underscoring its superior stability.

Cells highlighted as 1st / 2nd / 3rd per column.

Method	Mean	Med.	Tri.	B-25%	W-25%
WP [27]	5.99	4.37	4.72	0.89	14.08
GW [7]	6.78	5.53	5.69	1.73	14.08
GE [47]	4.73	3.38	3.80	1.11	10.52
SoG [15]	5.55	4.50	4.74	1.05	11.84
LSRS [16]	6.00	5.02	5.10	1.46	12.52
GP [51]	4.56	3.51	3.84	0.82	10.01
GI [36]	4.09	2.81	3.17	0.67	9.65
FC⁴ [20]	2.48	1.79	2.05	0.55	5.57
C⁴ [52]	2.23	1.49	1.75	0.51	5.05
C⁵ [1]	2.18	1.55	1.78	0.52	4.95
PCC [53]	2.39	1.60	1.76	0.51	5.47
ePCC [31]	2.16	1.39	1.61	0.45	5.08
SAFE w/o LCS	2.12	1.44	1.56	0.43	5.01
SAFE w/ LCS (P1)	2.03	1.38	1.50	0.39	4.80
SAFE w/ LCS (P4)	1.94	1.27	1.39	0.36	4.66

Table 2: Evaluation results on the Gehler-Shi dataset. We report the mean angular error (in degrees) based on 3-fold cross-validation. The results prove that SAFE is not limited to pure-color scenes, but achieves comparable performance on general scenes.

Cells highlighted as 1st / 2nd / 3rd per column.

Method	Mean	Med.	Tri.	B-25%	W-25%
WP [27]	7.55	5.68	6.35	1.45	16.12
GW [7]	6.36	6.28	6.28	2.33	10.58
SoG [15]	4.93	4.01	4.23	1.14	10.20
PCA-CC [10]	3.52	2.14	2.47	0.50	8.74
Woo et al. [50]	4.30	2.86	3.31	0.71	10.14
GI [36]	3.07	1.87	2.16	0.43	7.62
C⁴ [52]	1.35	0.88	0.99	0.28	3.21
CLCC [32]	1.44	0.92	1.04	0.27	3.48
GCC [8]	1.91	1.80	1.84	0.60	3.46
PCC [53]	2.64	1.61	1.83	0.45	6.62
ePCC [31]	2.45	1.59	1.78	0.44	5.87
SAFE w/o LCS	2.35	1.69	1.78	0.47	5.57
SAFE w/ LCS (P1)	2.33	1.66	1.75	0.47	5.43
SAFE w/ LCS (P4)	2.23	1.56	1.69	0.41	5.24

Qualitative comparison on images from the PolyU Pure Color dataset V2. Images are gamma-corrected for visualization. The angular error (in degree) is reported in the bottom-right corner of each image.

$**Scene-dependent LCS deviation versus scene entropy.** (Left figure) The LOWESS smoother with $95\%$ bootstrap confidence band. {(Right figure)} Hexbin density of the same data with the LOWESS smoother overlaid in red. Quantitatively, the lowest-entropy quintile ($n{=}255$) attains $\bar{\Delta}_w = 0.0547$ versus $0.0462$ for the highest-entropy quintile ($18\%$ relative increase on pure-color scenes) and the Pearson correlation $r(H,\Delta_w) = -0.149$ is highly significant ($p < 10^{-7}$, $N{=}1271$). The analysis validates that the LCS design further improves the accuracy of illumination estimation in pure-color scenes.$

Scene-dependent LCS deviation versus scene entropy. (Left figure) The LOWESS smoother with $95%$ bootstrap confidence band. {(Right figure)} Hexbin density of the same data with the LOWESS smoother overlaid in red. Quantitatively, the lowest-entropy quintile ($n{=}255$) attains $\bar{\Delta}_w = 0.0547$ versus $0.0462$ for the highest-entropy quintile ($18%$ relative increase on pure-color scenes) and the Pearson correlation $r(H,\Delta_w) = -0.149$ is highly significant ($p < 10^{-7}$, $N{=}1271$). The analysis validates that the LCS design further improves the accuracy of illumination estimation in pure-color scenes.

$**Introspection on the SAFE backbone.** The modulation attention $\alpha(\mathbf{g})$ quantifies how strongly the modulation network preserves each feature dimension; the input projection weight $\lVert W_0[:,j]\rVert_2$ quantifies how much linear weight the backbone allocates to each dimension; and the intrinsic effective contribution quantifies how important each token is to the illumination estimation.$

Introspection on the SAFE backbone. The modulation attention $\alpha(\mathbf{g})$ quantifies how strongly the modulation network preserves each feature dimension; the input projection weight $\lVert W_0[:,j]\rVert_2$ quantifies how much linear weight the backbone allocates to each dimension; and the intrinsic effective contribution quantifies how important each token is to the illumination estimation.

BibTeX

@article{lee2026safe,
  title={SAFE: Scene-Aware Feature Modulation for Color Constancy with Learned Color Sapce in Pure-Color Scenes},
  author={Lee, Yuan-Kang and Chen, Kuan-Lin and Chang, Chih-Heng and Ding, Jian-Jiun},
  journal={arXiv preprint arXiv:2601.05249},
  year={2026}
}