基于球面谐波系数的室内混合现实场景光照估计OA

In mixed reality applications,recovering a high dynamic range(HDR)panoramic lighting map from a single low dynamic range(LDR)image with a limited field of view is a critical challenge for achieving realistic scene fusion.To address this issue,this paper proposes RGLight,an indoor scene lighting estimation method based on spherical harmonic(SH)coefficients.Firstly,parallel convolution paths and gated multilayer perceptrons(MLPs)are added to improve the Vision Transformer for feature extraction from limited-view images,regressing SH coefficients that approximate scene lighting to re-cover low-frequency lighting.Next,a color constancy loss is introduced to reduce the color discrepancy between the extracted lighting information and the ground-truth lighting during the lighting parameter regression stage.Then,the SH coefficients are used to reconstruct SH irradiance maps to guide the gen-eration of HDR panoramic lighting maps,restoring complete lighting information.Finally,the render-ing results of generated images and real images are compared to validate the proposed method.Theoreti-cal analysis and experimental results show that RGLight achieves excellent performance in both quanti-tative and qualitative evaluations.For diffuse gray and matte silver material latitude-longitude spheres rendering using the panoramic lighting map predicted by RGLight,the root mean square errors relative to ground-truth lighting rendering are 0.050 9 and 0.052 1,respectively,and the angular errors are 4.903 7° and 5.172 2°,respectively,validate the effectiveness and accuracy of the proposed method.