Leveraging GPU advances for transparency, compositing and lighting using deep images and light grids

<p>This thesis explores data structures built by leveraging recent GPU features combined with fast rasterization, specifically deep images and light grids. The primary goal of computer graphics is rendering better images faster. There are many rendering approaches, of which rasterization offers fast performance which is well suited to realtime computer graphics. Graphics processing units (GPUs) are primarily but not exclusively designed to accelerate rasterization which requires both object space and image space computations. For many problems image space computation can be advantageous for improving performance, accuracy and simplicity, a classic example being the z-buffer. By leveraging fast rasterization and other capabilities of modern GPUs, all rendered geometry can now be captured and stored in realtime as per-pixel lists of fragments in deep images, which allows solving more problems in image space than previously possible using regular 2D flat images. This gives rise to new image based rendering techniques as well as performant improvements to previous techniques.</p> <p>Deep images and light grids are investigated for the applications of deep compositing, transparency and lighting. Efficient ways of using GPU hardware are presented for both building and using the underlying data structures. Performance improvements are presented for both building and merging deep images, allowing realtime deep compositing and transparency of more complex scenes than previously possible, including to our knowledge, the first investigation of fast deep image merging approaches. A new approach for building a light grid is presented for realtime rendering of scenes with many lights, allowing more lights than previous techniques. Where deep images are typically a view-dependent discretized representation of geometry, an improved approach which captures multiple deep images simultaneously for building a view independent representation termed a compound deep image (CDI) is presented which is suitable for realtime raycasting. Realtime indirect lighting is explored combining both the CDI and light grid, adding to the ever expanding range of GPU accelerated image space techniques.</p>