Results
Middlebury Sequences without Ground Truths
In this experiment, we evaluate the visual quality of our algorithm on Middlebury datasets with
hidden ground truth. Consequently, we can not measure the errors of our
solutions, but they are available on the pages of the Middlebury benchmark. In
July 2010, our method ranks seventh out of 37, and represents the fastest
approach among the top 10 methods.
In order to stay comparable with the Middlebury benchmark, we use fixed
parameters this time. As in our last experiment,
we encode the flow field by a colour code indicating strength and direction of
the observed motion. For each sequence, this visualisation can be found at the
right. To the left, we show the two frames used to compute this flow field.
As before, the required run-time on a GeForce GTX 480 is depicted below the
images.
Please click on any of the images for a larger version.
Army, 584×388 px. Runtime: 653 (696) ms.
Mequon, 584×388 px. Runtime: 654 (696) ms.
Schefflera, 584×388 px. Runtime: 654 (696) ms.
Wooden, 584×388 px. Runtime: 654 (696) ms.
Grove, 640×480 px. Runtime: 735 (783) ms.
Urban, 640×480 px. Runtime: 734 (782) ms.
Yosemite, 316×252 px. Runtime: 463 (487) ms.
Teddy, 420×360 px. Runtime: 571 (602) ms.
Backyard, 640×480 px. Runtime: 734 (783) ms.
Basketball, 640×480 px. Runtime: 734 (782) ms.
Dumptruck, 640×480 px. Runtime: 733 (782) ms.
Evergreen, 640×480 px. Runtime: 734 (783) ms.
< Results with GT
Main Page
Runtimes >
|