GPU-based Deep Image Post Rendering And Compositing System

We build a GPU-based deep image rendering and compositing system that integrates multiple deep image post-processing techniques and rendering algorithms. Comparing to or-dinary images, deep images encode more data in a pixel, which introduces memory problems for the processing system. To solve this, we use adaptive transparency (AT) buffer to com-press deep images on demand to make intermediate memory usage predictable. Using AT buffer also increases the post-processing performance. We also present a deep image space ray tracing algorithm and an adaptive temporal sampling method to simulate high quality DOF and realtime post-processed motion blur, respectively. Meanwhile, to further exploit the depth information in deep images, we implement fog effects with3D noise and light beams in image space and composite them with deep images. All the image processing algo-rithms in our system are implemented on the GPU, achieving computing speed-ups ranging from20to300times over the CPU counter part. The speed-up enables realtime feedback. Our system design is based on the node graph model and consists of an interaction module and an execution module. The user can combine multiple pre-defined nodes to achieve the desired post-processing effect. Our system ensures nodes are applied in an optimized way and script-based procedures are called properly.