PRefToMotion

C++ plugin for Foundry's Nuke software that calculates the motion vectors from a source frame to the current frame using a position reference pass or similar. It uses the nanoflann v1.8 kd-tree library under the hood to do a nearest neighbour lookup using the xyz value of the supplied pref channel.

example image a smiley face drawn at frame 1 and warped to 5, 10 and 15 using a position reference pass

usage

PRefToMotion(.dylib|.so|.dll) should now be built and installed into your user's .nuke folder, so now it's just a matter of launching nuke and opening the test script pref_motion_example.nk

parameters

Very simple list of parameters to interact with:

• pref channels: channels that hold the xyz pref or similar data. If a 4th channel is supplied, it will be used as a mask.
• uv channels: channels to store the uv data that is being generated
• mask: optional mask channel
• source frame: source frame to calculate the motion vectors from
• samples: number of nearest neighbours to look up, used to calculate a weighted average from the squared distance to the lookup point
• mode: generates the motion as either an st map (normalised), uv map (vectors) or as the source pixels

Performance

• build threads: number of threads to use for building the kd-tree (0 = auto)
• use rknn: enable radius-based k-nearest neighbor search for improved performance
• max radius: maximum search radius for rknn queries (only used when "use rknn" is enabled). Smaller values can significantly improve query speed by limiting the search space.

speed

Building the indices and querying a kd-tree with more than a million points can be slow. The new rknn (radius k-nearest neighbor) feature with max radius parameter provides significant performance improvements by limiting the search space during queries.

example timing of building of a kd-tree with 1,063,379 points

--------------- kd tree ---------------
           bbox: 91, 48, 1469, 1023
output channels: rgba,forward.u,forward.v
  pref channels: rgb
     num points: 1063379
        samples: 3
   source frame: 1
  current frame: 15
--------------- timing ---------------
     query time: 208 ms
     build time: 1096 ms
     total time: 1304 ms

Performance can be further improved by:

• Enabling use rknn checkbox
• Tuning the max radius parameter (smaller values = faster queries, but may miss distant neighbors)
• Adjusting build threads for faster kd-tree construction

dependencies

• nanoflann v1.8 - A C++11 header-only library for building KD-Trees