Kavita Bala

Rendering and modeling complex scenes is challenging. Understanding and exploiting how humans perceive complex scenes is an important area in graphics. We have worked on multiple projects in this area.

Understanding how we perceive complex geometric aggregates is an open problem. We study the perception of aggregates to derive metrics for scene simplification (SIG '08) (Project).

Rendering high complexity scenes including complex illumination and rendering effects such as motion blur, participating media, global illumination, and depth-of-field, is challenging. Multidimensional lightcuts (SIG '06) and lightcuts (SIG '05) present a unified, scalable rendering framework to efficiently render complex scenes with such effects. By unifying complex illumination into one framework we achieve high scalability and accurate imagery.

(Multidimensional Lightcuts Project, Subsurface Lightcuts Project, Lightcuts Project).

Previewing still images and animations of scenes with high geometric and illumination complexity, and arbitrary shading models, is useful for applications such as cinematic lighting design. Matrix row-column sampling (SIG '07) treats rendering as the evaluation of a very large matrix of pixel-light interactions; this matrix can be efficiently approximated by evaluating a very small set of pixels, and using them to cluster lights globally, for a fast approximation of the image (Project).

Tensor clustering extends this idea to render animations including deforming characters. This work extends the row-column sampling approach to tensors, and introduces a clustering metric that minimizes temporal flicker (Project).

When a user changes the scene (but not the lighting), rapidly identifying the parts of the radiance computation that are affected by user manipulation is difficult. 5D Ray Segment Trees (EGRW '99) efficiently identify affected radiance interpolants and incrementally ray trace images. (Project).

The human visual system is sensitive to features such as silhouettes and shadows.

Edge-and-point rendering (SIG '03) identifies visually important features (edges) and combines them with sparse, expensive shading samples to achieve interactive rendering with global illumination. This approach bridges the gap between expensive, high-quality rendering and fast, interactive display. Project, GPU implementation project (GI '06)

Feature-based textures (EGSR '04) are a resolution-independent representation of textures that capture visually important features. FBT Project

Constrained Minimization Synthesis (TVCG '06) casts detail synthesis and image analogies as an energy minimization problem, and uses graph cut techniques to synthesize textures while satisfying constraints.
Project

Local illumination environments (EGSR '02) capture the part of the environment that influences shading at each part of a scene. This approach enables rendering with complex direct illumination including hundreds of lights. LIE Project

Iterative adaptive sampling (TVCG '06) efficiently renders scenes with many lights by adapting the sampling distribution of the lights in a multipass algorithm.