Lightmap baker

A tool to accelerate calculation of Ambient Occlusion for complex scenes on the GPU. Done as a master thesis in 2007/2008.

It is now part of the 3D city visualization tool LandXplorer by Autodesk. Calculating the light maps takes between a few seconds and a few minutes depending on the model size. (12s to generate 14 1024^2 textures for a 15k triangle model using 400 light sources on a Geforce 8600M 128MB notebook graphics card with Core Duo 2.2Ghz processor).

The algorithm consists of the following steps:

splitting city model into chunks to support large models
identifying mesh connectivity
clustering of triangles to optimize for lowest texture size
generating UV parameterization
distributing virtual light sources to mimic realistic outdoor lighting
accumulation of light's contributions using graphics hardware

See the images for more details.

A small city model rendered with the baked lighting. The white buildings do not have textures in the model. Highlighting soft shadows. Note the sneaky artifact in the lower left corner caused by shadow map settings not being set optimally for the scene. The same model rendered without any textures. This makes it easier to evaluate the light map quality and spot artifacts. A slightly larger city model. Controlling the number of virtual lights is an easy way to make trade offs between quality and rendering time. The artifacts caused by having too few of them can be seen on the shadows of some of the higher buildings. A visualization of the generated texture atlas. As a first step connectivity information is generated by creating an indexed mesh using a spatial hash. Then clusters are formed with the goal of forming rectangular-ish shapes when unwrapped onto a 2d plane. A hill climbing algorithm adds them by adding triangles from the clusters frontier that maximize utilization of the cluster's AABB. Then clusters are distributed onto a set of 2d textures. They get sorted by width and then placed left-to-right and top-to-bottom. When a texture is full unused areas at the end of each strip of clusters will be collected and later filled using smaller clusters. By supporting colored virtual point lights simple image based lighting (single bounce) can be achieved.