Benchmark results for a simple rotation transform, various interpolations

Parent ↗

Summary

• The interpolation methods falls into three groups.

  • near-neighbour and bilinear as single fast group, followed by
  • bicubic as much slower, and
  • lanczos a few orders slower than that.

• Correlates very well with the size of the neighbourhood used to compute each interpolated value from.

• Todo

  • See how much this is influenced by the separate fetching of the necessary neighbourhoods for each result pixel.
  • This requires the writing of an alternate implementation attempting to fetch a larger region useful to computing multiple results. I.e. a bit of an operator-local cache.
  • Also, there might be some threshold where individual fetching is more advantegous, because it would not incur the overhead of computing much of the input image.

  • It is unclear at this point if the threshold is better as

    • a percentage of the input we wish to stay below of, or as
    • a ratio between result and fetched area, or as
    • a ratio between the area of the union of the separate regions vs the area of the single region fetched

Plots