Adaptive rendering using Markov model

To address the realistic problems of current sampling and reconstruction methods, this paper proposes an adaptive rendering algorithm based on the Markov model and feature information. First, the rendering space is coarsely ray-traced and a one-step state transition probability matrix is built for each pixel. Second, per-pixel error is estimated by computing the multi-step transition probability matrix, which is coupled with the one-step matrix to determine the noise level of each pixel. In the reconstruction stage, the estimated pixel error is used to select a suitable scale of cross-bilateral filter, which make a trade-off between robustness to noise and fidelity to image details. In addition, feature buffers of texture, depth and surface normal are considered to preserve scene details. The experimental results demonstrate that our algorithm outperforms previous approaches both in visual image quality and numerical error.