Filtering-Based Reconstruction for Gradient-Domain Rendering
DescriptionGradient-domain rendering methods reconstruct color images based on the Poisson equation with gradients from correlated sampling. The relatively low variance in the gradient estimation facilitates convergence but the inevitable noises make the solving process prone to unpleasant spiky artifacts.
We present a gradient-guided filtering approach for reconstruction, which avoids the instability from the direct usage of noisy gradients. Instead, we model the output color of each pixel as a weighted combination of neighboring pixels, where the gradients are used as guidance to compute optimized filtering weights. The gradients are enhanced before being used in gradient-guided filtering. A coarse-to-fine strategy is also employed to make use of information from a larger scale.
Experiments demonstrate that our method achieves the best reconstruction results for gradient-domain renderings compared to existing techniques. Besides, our method has two desirable properties: first, our method is not learning-based so it does not require an extra training step and would be more robust for unseen scenes; second, our method is designed to be asymptotic unbiased.
We present a gradient-guided filtering approach for reconstruction, which avoids the instability from the direct usage of noisy gradients. Instead, we model the output color of each pixel as a weighted combination of neighboring pixels, where the gradients are used as guidance to compute optimized filtering weights. The gradients are enhanced before being used in gradient-guided filtering. A coarse-to-fine strategy is also employed to make use of information from a larger scale.
Experiments demonstrate that our method achieves the best reconstruction results for gradient-domain renderings compared to existing techniques. Besides, our method has two desirable properties: first, our method is not learning-based so it does not require an extra training step and would be more robust for unseen scenes; second, our method is designed to be asymptotic unbiased.
Event Type
Technical Papers
TimeThursday, 5 December 20249:28am - 9:42am JST
LocationHall B5 (2), B Block, Level 5
