Occlusion Culling on the GPU : Inner Conservative Occluder Rasterization

Abstract: Context. Many occlusion culling algorithms have to cope with the task of balancing performance and accuracy. While it is desirable to accurately identify all occluded scene objects, settling with a rough estimate is often more beneficial for the overall performance. Algorithms that rely on a depth buffer can often gain a lot of performance by performing the occlusion culling at a lower resolution than the resolution of the screen. This calls for more advanced methods to render the depth buffer as the standard rasterizer will not guarantee inner coverage. Objectives. The goal of this thesis is to find a solution to generate a depth buffer where all rasterized pixels are fully covered by overlapping occluders. An algorithm is proposed that is based on previous work on inner conservative rasterization. The algorithm addresses some of the problems existing methods are suffering from, but also has some flaws of its own. Methods. The proposed algorithm is tested by comparing it to two methods that also produce conservative results. A GPU-based occlusion culling system is developed to conduct an experiment. The experiment is performed by measuring performance and culling efficiency in two different views of a scene. The scene is set up to represent an average setting in a game. Results. The results from the experiment show that the proposed algorithm outperforms its competitors in many cases. In the first scene view, the total frame time is 5% faster at a full screen resolution of 1366x768 pixels and 8% faster at a full screen resolution of 1920x1080 pixels. The depth buffer generated by the proposed algorithm is culling atleast as many occludees as its competitors and often surpasses them. In the second scene view, the total frame time is 2% faster at a full screen resolution of 1366x768 pixels and 3% faster at a full screen resolution of 1920x1080 pixels. The depth buffer generated by the proposed algorithm is often culling more occludees than its competitors, but is at lower resolutions less efficient, up to 3%. Conclusions. The conclusions show that the goal has been reached. The proposed algorithm lacks flexibility, but provides good performance and accuracy. Future work to improve the proposed algorithm is outlined.