Bridging Intrinsic Image Decomposition and Color Constancy for Zero-Shot Single and Multi-Illuminant Estimation

Computational Color Constancy (CC) aims to estimate and correct scene illumination in order to recover consistent surface colors under varying lighting conditions. In parallel, intrinsic image decomposition (IID) separates images into albedo and shading components, providing structured cues about scene illumination. Despite their close relationship, these two tasks are typically addressed independently as the domains of CC and IID differ considerably, partly due to sensor-dependent spectral sensitivity functions.

In this work, we establish a connection between IID and CC by introducing a zero-shot, cross-camera framework that repurposes a pre-trained IID model for CC without any task-specific training. We propose a novel illumination estimation pipeline that explicitly bridges the domain gap between IID and CC. The input image is pre-balanced using a statistical CC method, after which the albedo is estimated, the shading chromaticity is implied, and a dense illumination map is obtained through chromaticity clustering and blending. This formulation naturally handles both single- and multi-illuminant scenes while remaining entirely independent of CC supervision.

In a zero-shot setting, the proposed method achieves, using the spatially varying illumination map, a mean angular error of 3.02° on Gehler-Shi, 2.69° on NUS-8, 3.21° on INTEL-TAU, and 2.33° on LSMI.

Our method achieves competitive performance and outperforms some of the cross-camera approaches in cross-camera/dataset evaluation protocols. Ablation studies show that only applying IID to CC substantially degrades performance, confirming the presence of a significant domain gap and demonstrating the necessity of the proposed bridging pipeline. These results establish IID as a powerful foundation for CC, enabling a sensor-agnostic framework for both single and, more importantly, multi-illuminant CC.

Runtime analysis shows that estimating a dense illumination map from a 512×512 image takes approximately 1–2 seconds on a single GPU, providing a practical baseline for future acceleration through model compression and optimization.