Abstract Correctly compositing transparent fragments is an important and long-standing open problem in real-time computer graphics. Multifragment rendering is considered a key solution to providing high-quality order-independent transparency at interactive frame rates. To achieve that, practical implementations severely constrain the overall memory budget by adopting bounded fragment configurations such as the k-buffer. Relying on an iterative trial-and-error procedure, however, where the value of k is manually configured per case scenario, can inevitably result in bad memory utilization and view-dependent artifacts. To this end, we introduce a novel intelligent k-buffer approach that performs a non-uniform per pixel fragment allocation guided by a deep learning prediction mechanism. A hybrid scheme is further employed to facilitate the approximate blending of non-significant (remaining) fragments and thus contribute to a better overall final color estimation. An experimental evaluation substantiates that our method outperforms previous approaches when evaluating transparency in various high-depth-complexity scenes.
展开▼
