Facoltà di scienze informatiche

Efficient multi-bounce lightmap creation using GPU forward mapping

Schärfig, Randolf ; Hormann, Kai (Dir.) ; Stamminger, Marc (Codir.)

Thèse de doctorat : Università della Svizzera italiana, 2016 ; 2016INFO008.

Computer graphics can nowadays produce images in realtime that are hard to distinguish from photos of a real scene. One of the most important aspects to achieve this is the interaction of light with materials in the virtual scene. The lighting computation can be separated in two different parts. The first part is concerned with the direct illumination that is applied to all surfaces lit by a... Plus

Ajouter à la liste personnelle
    Summary
    Computer graphics can nowadays produce images in realtime that are hard to distinguish from photos of a real scene. One of the most important aspects to achieve this is the interaction of light with materials in the virtual scene. The lighting computation can be separated in two different parts. The first part is concerned with the direct illumination that is applied to all surfaces lit by a light source; algorithms related to this have been greatly improved over the last decades and together with the improvements of the graphics hardware can now produce realistic effects. The second aspect is about the indirect illumination which describes the multiple reflections of light from each surface. In reality, light that hits a surface is never fully absorbed, but instead reflected back into the scene. And even this reflected light is then reflected again and again until its energy is depleted. These multiple reflections make indirect illumination very computationally expensive. The first problem regarding indirect illumination is therefore, how it can be simplified to compute it faster. Another question concerning indirect illumination is, where to compute it. It can either be computed in the fixed image that is created when rendering the scene or it can be stored in a light map. The drawback of the first approach is, that the results need to be recomputed for every frame in which the camera changed. The second approach, on the other hand, is already used for a long time. Once a static scene has been set up, the lighting situation is computed regardless of the time it takes and the result is then stored into a light map. This is a texture atlas for the scene in which each surface point in the virtual scene has exactly one surface point in the 2D texture atlas. When displaying the scene with this approach, the indirect illumination does not need to be recomputed, but is simply sampled from the light map. The main contribution of this thesis is the development of a technique that computes the indirect illumination solution for a scene at interactive rates and stores the result into a light atlas for visualizing it. To achieve this, we overcome two main obstacles. First, we need to be able to quickly project data from any given camera configuration into the parts of the texture that are currently used for visualizing the 3D scene. Since our approach for computing and storing indirect illumination requires a huge amount of these projections, it needs to be as fast as possible. Therefore, we introduce a technique that does this projection entirely on the graphics card with a single draw call. Second, the reflections of light into the scene need to be computed quickly. Therefore, we separate the computation into two steps, one that quickly approximates the spreading of the light into the scene and a second one that computes the visually smooth final result using the aforementioned projection technique. The final technique computes the indirect illumination at interactive rates even for big scenes. It is furthermore very flexible to let the user choose between high quality results or fast computations. This allows the method to be used for quickly editing the lighting situation with high speed previews and then computing the final result in perfect quality at still interactive rates. The technique introduced for projecting data into the texture atlas is in itself highly flexible and also allows for fast painting onto objects and projecting data onto it, considering all perspective distortions and self-occlusions.