Test images

For this benchmark we've chosen 14 2D images of size 256x256 or 512x512. They were chosen in such a way to be representative of usual applications or to verify how some functions behave. As an example :

• lena and astronaut are standard gray level images. With bubbles_gray, these are interesting images to benchmark segmentation;

• bubbles_gray image come from material sciences applications;

• tools is an interesting image when trying to isolate objects in gray level images (segmentation);

• bubbles_bin, eutectic hubble_EDF and balls are images interesting to experiment label() and areaThreshold();

• balls, eutectic and hubble_bin images are binary images with very different numbers of regions.

These images were scaled up and down to cover the range from 256x256 to 8192x8192. This procedure allows us to check how algorithms behave with different image sizes, with the same complexity.

All images were submitted to all functions even if sometimes it makes no sense to apply some function to some specific image. As an example, segmentation of binary images is useful only to separate superposing objects (images coffee and cells) - anyway, the function was applied to all binary images.

It's also important to benchmark some functions (e.g. segmentation) with respect to their complexity. To do this, one usual and practical procedure is to create a mosaic of smaller images with varying repetition factors. This will be integrated in some future version of this benchmark.

Neverthless, a mosaic of small images was created to investigate CPU and memory usage of both libraries (see resources usage on taurus and resources usage on nestor).

Binary images

alumine.png	balls.png	bubbles_bin.png
256 x 256	256 x 256	256 x 256
cells.png	coffe.png	eutectic.png
256 x 256	256 x 256	256 x 256
gruyere.png	hubble_EDF_bin.png	metal.png
256 x 256	512 x 512	256 x 256

Gray images

astronaut.png	bubbles.png	lena.png
512 x 512	512 x 512	256 x 256
hubble_EDF_gray.png	tools.png
512 x 512	256 x 256