Object Color Relighting with Progressively Decreasing Information

Abstract Object color relighting, the process of predicting an object’s colorimetric values under new lighting conditions, is a significant challenge in computational imaging and graphics. This technique has important applications in augmented reality, digital heritage, and e-commerce. In this paper, we address object color relighting under progressively decreasing information settings, ranging from full spectral knowledge to tristimulus-only input. Our framework systematically compares physics-based rendering, spectral reconstruction, and colorimetric mapping techniques across varying data regimes. Experiments span five benchmark reflectance datasets and eleven standard illuminants, with relighting accuracy assessed via △ E 00 metric. Results indicate that third-order polynomial regressions give good results when trained with small datasets, while neural spectral reconstruction achieves superior performance with large-scale training. Spectral methods also exhibit higher robustness to illuminant variability, emphasizing the value of intermediate spectral estimation in practical relighting scenarios.