Complementary Optic Flow on GPU - Mathematical Image Analysis Group, Saarland University

Complementary Optic Flow on the GPU

Results
Middlebury Sequences with Ground Truths

In this experiment, we evaluated the results of our algorithm w.r.t. the Average Angular Error (AAE) and the Average Endpoint Error (AEE) using the Middlebury datasets with known ground truths. For all sequences, we optimised the parameters of our algorithm individually, such that the results reflect the the best quality that can be obtained with our approach. In each case, we depict the left frame of the input data on the left side, followed by the colour-coded ground truth in the center, and our result to the right. The visualisation scheme we use encodes relative intensities by brightness, and directions by colours, as depicted by the colour circle next to this text. White areas denote regions where no ground truth data is available.

In addition, we measured the run-times required on a Sparkle GeForce GTX 480. Depicted is the pure run-time of the algorithm, as well as the overall run-time including up- and download of the problem between CPU and GPU (in parantheses). Note that this GPU is slightly more powerful than the one we used in the experiments for our paper, and the run-times are thus smaller.

Please click on any of the images for a larger version.

Dimetrodon, 584×388 px. AAE: 1.49°, AEE: 0.08. Runtime: 226 (264) ms.

Grove2, 640×480 px. AAE: 2.32°, AEE: 0.16. Runtime: 366 (410) ms.

Grove3, 640×480 px. AAE: 5.30°, AEE: 0.53. Runtime: 365 (411) ms.

Hydrangea, 584×388 px. AAE: 2.01°, AEE: 0.17. Runtime: 322 (361) ms.

RubberWhale, 584×388 px. AAE: 2.93°, AEE: 0.09. Runtime: 322 (361) ms.

Urban2, 640×480 px. AAE: 2.75°, AEE: 0.29. Runtime: 725 (774) ms.

Urban3, 640×480 px. AAE: 3.52°, AEE: 0.39. Runtime: 725 (774) ms.

Venus, 420×380 px. AAE: 3.99°, AEE: 0.28. Runtime: 274 (302) ms.

< FED Main Page Results without GT >