Physically‐Constrained Diagonal Manifolds for Cross‐Sensor Color Constancy

Cross‐sensor color constancy is affected by variations in sensor spectral sensitivity, often leading models trained on one sensor to degrade when applied to another. To address this challenge, we propose a physically motivated training strategy that improves cross‐sensor robustness without requiring explicit sensor characterization or multisensor training data. The proposed method incorporates the von Kries diagonal model as a dynamic training constraint. Specifically, illuminants are projected onto a diagonal manifold derived from measured ground‐truth illuminants, thereby simulating structured variations in the image formation process. This transformation preserves consistency with the radiometric model while encouraging the network to disentangle illuminant‐related factors from sensor‐dependent characteristics. As a result, a model trained on a single sensor can generalize effectively to unseen sensors. Experimental results on Cube+, INTEL‐TAU, and NUS‐8 demonstrate the effectiveness of the proposed strategy.