Key Point Subspace Acceleration and soft caching

Many applications in Computer Graphics contain computationally expensive calculations. These calculations are often performed at many points to produce a full solution, even though the subspace of reasonable solutions may be of a relatively low dimension. The calculation of facial articulation and rendering of scenes with global illumination are two example applications that require these sort of computations. In this paper, we present Key Point Subspace Acceleration and Soft Caching, a technique for accelerating these types of computations. Key Point Subspace Acceleration (KPSA) is a statistical acceleration scheme that uses examples to compute a statistical subspace and a set of characteristic key points. The full calculation is then computed only at these key points and these points are used to provide a subspace based estimate of the entire calculation. The soft caching process is an extension to the KPSA technique where the key points are also used to provide a confidence estimate for the KPSA result. In cases with high anticipated error the calculation will then "fail through" to a full evaluation of all points (a cache miss), while frames with low error can use the accelerated statistical evaluation (a cache hit).