[Openexr-devel] Interpreting Deep Images, Revised Document

[Florian Kainz]

To the OpenEXR users and developers who have read the previous version
of my paper about interpreting deep image data:

Attached is a revised version of the document.  Most of the text has been
re-written from scratch based on extensive feedback from (in alphabetical
order) Rod Bogart at Pixar, Ed Hanway at ILM, Peter Hillman at Weta Digital,
Christopher Horvath at Pixar, Piotr Stanczyk at ILM and Jon Wadelton at
The Foundry.

In terms of content, the most important changes are:

        - The math for merging overlapping volume and point samples was
          incorrect in the previous version.  This has been fixed.

        - The document includes numerically stable C++ code for splitting
          a sample at a given depth and for merging two overlapping samples.
          The code properly handles samples with very small alphas.

        - Samples in a single pixel are explicitly allowed to be unsorted
          and overlapping.  The previous version assumed that samples always
          had to be sorted and non-overlapping, but that is not how deep
          images are used in practice.

        - Merging two images, and making the samples in an image sorted and
          non-overlapping, are now described as separate operations.

        - The document describes a method for finding the alpha channel that
          is associated with a given color channel in a multi-layer image.

        - Deep images are required to have exactly one or two depth channels.

I expect current version of the document to become part of the official
OpenEXR documentation, and I hope Peter Hillman will soon publish a paper
that explains the math for merging overlapping samples.

Florian

Interpreting OpenEXR Deep Pixels-10-25-13.pdf
Description: Adobe PDF document