Expand description
Image comparison (diffing) for computer graphics renderer test cases.
The algorithm implemented in this library is intended to allow comparing images of the same scene which were rendered using different algorithms, or different hardware, causing small “rounding errors” in either color or spatial position.
Principle of operation
Suppose we are comparing two images, A and B. For each pixel in A (except for the perimeter), a neighborhood around the corresponding pixel in B is compared, and the smallest color difference is taken to be the difference value for that pixel in A. Then, the same process is repeated, swapping the roles of the two images, and the final difference value for each pixel is the maximum of those two.
The pixel differences are then compiled into a difference image (for user viewing) and a histogram (for pass/fail conditions).
The effect of this strategy is that any feature in the image, such as the edge of a shape, can be displaced by up to the neighborhood size in any direction, thus tolerating different choices of rounding into the pixel grid, as long as the color is the same.
This algorithm does not inherently accept differences in antialiased images, because depending on how an edge lands with respect to the pixel grid, the color may be different. A future version of this library may solve that problem by accepting any color which is a blend of colors found in the neighborhood.
Structs
- Output of
diff(), a comparison between two images. - A histogram of color difference values.
- A bound upon pixel differences observed in a
Difference.
Functions
- Compares two images with a neighborhood-sensitive comparison which counts one pixel worth of displacement as not a difference.