Fast Reconstruction for Monte Carlo Rendering Using Deep Convolutional Networks

Denoising the Monte Carlo (MC) rendering images is different from denoising the natural images since low-sampled MC renderings have a higher noise level and there are inexpensive by-products (e.g., feature buffers) we can leverage. However, the main challenge is designing a model to fast fuse these feature buffers and reconstruct perceptually noise-free images from noisy MC renderings. In recent years, supervised learning methods remove the noise and reconstruct clean images, but most of them cannot handle MC noise well. In this paper, we introduce an end-to-end CNN model to fuse feature buffers and predict a clean image directly. In addition, we devise a new high-dynamic range (HDR) image normalization method to help us to train the model on HDR images in a more efficient and stable way. We setup a series of experiments for selecting the hyperparameter of our deep learning model, network depth, which can promote our network's performance and avoid overfitting. We demonstrate that our model is robust on a wide range of scenes and can generate satisfactory results in a significantly faster way.