Geometry-aware Compressive Dictionary Learning based Rendering

Compressive Sensing (CS) is a new sampling theory. It states that we can reconstruct a signal from very few measurements taken by projecting the signal rather than point sampling it. The signal can be reconstructed if it is sparse or sparse in some domain. This theory was employed recently in [1] to accelerate the rendering of ray-traced images, by rendering just a subset of pixels then applying the CS reconstruction to fill the missing ones using wavelet as a transform domain to seek the signal sparsity condition. In this paper, we use a learned dictionary rather than standard wavelet to better sparsify our images and hence improve the CS reconstruction. We also inject cheap geometry information (depth) to accurately reconstruct our images. Finally, we post-process our images by applying a modified version of the bilateral filter to improve the overall quality. Obtained results show a clear improvement in the quality of the image reconstruction while accelerating the rendering time as compared to [1].