Motion Blur Ray Tracer

Truncated box (top) vs. tent (bottom) time filters

From Fall 2010 until May 2011, I worked on my M.S. thesis project, which was to explore the perceptual impact of various motion blur techniques in distribution ray tracing. My advisor for this project was Prof. Bobby Bodenheimer, and in this work we hoped to achieve a deeper understanding into the best methods for efficiently producing high-quality motion blur. To accomplish this goal we conducted a suite of user trials in order to measure perceptual quality in a variety of configurations. We used the method of constant stimuli to find the just noticeable difference of noise with respect to the different configurations, and the aspects of motion blur that we investigated included:

  • Distribution of samples across time (triangle filter, truncated box filter)
  • Number of samples per pixel
  • Point set used for sample generation (Hammersley, Poisson disk, stratified monte-carlo)

In order to generate the large number of test videos that we needed, I wrote a custom renderer for this project using OptiX that allows for independent adjustment of these various parameters. In addition, the renderer supports the following features:

  • OBJ meshes and materials
  • Instancing
  • Grouping (for controlling motion en masse)
  • Rotational and translational motion paths for objects, groups, and camera
  • BVH acceleration structure with efficient bounding box calculation
  • Filter importance sampling for (u,v,t) coordinates of samples
  • Batch rendering

Rotating camera and spheres

Texture-mapped spheres in motion with three different camera shutter times

The appearance of motion blur with three different point sets and three sampling levels

The application of filter importance sampling. 125-point Hammersley set unmodified (left) and Gaussian-warped (right)

Measured psychometric functions for both experimental scenes. The red line represents the JND cutoff, and the green line represents random chance.

Measured psychometric functions for all three experimental point sets. The red line represents the JND cutoff, and the green line represents random chance.