Uncertainty Visualization by Representative Sampling from Prediction Ensembles
Sampling is a core process for a variety of graphics applications. Among existing sampling methods, blue noise sampling remains popular thanks to its spatial uniformity and absence of aliasing artifacts. However, research so far has been mainly focused on blue noise sampling with a single class of samples. This could be insufficient for common natural as well as man-made phenomena requiring multiple classes of samples, such as object placement, imaging sensors, and stippling patterns. We extend blue noise sampling to multiple classes where each individual class as well as their unions exhibit blue noise characteristics. We propose two flavors of algorithms to generate such multi-class blue noise samples, one extended from traditional Poisson <i>hard</i> disk sampling for explicit control of sample spacing, and another based on our <i>soft</i> disk sampling for explicit control of sample count. Our algorithms support uniform and adaptive sampling, and are applicable to both discrete and continuous sample space in arbitrary dimensions. We study characteristics of samples generated by our methods, and demonstrate applications in object placement, sensor layout, and color stippling.