In the paper Scene flow estimation using Intelligent Cost Functions we showed that many common motion estimation cost functions are unable to distinguish correct and incorrect motion estimates. We also showed that machine learning techniques could be employed to learn a more accurate, non-parametric cost function, which significantly improved motion estimation accuracy when integrated into existing scene flow and optical flow algorithms.
Supplementary material
To accompany the results in the paper, the analysis was performed on a wide range of additional scenes to show the generality of the findings. The ICFs used to generate these results were trained on unseen data (for example, ICFs tested on Kitti sequences were trained on greyscale Middlebury scenes). Both the ICFs and the standard motion estimation metrics, are evaluated with 3 different norms.
As discussed in the paper, the type of norm used has no noticable effect on the performance of the metrics. In addition, regardless of the scene, the conclusions drawn about the individual metrics remain the same (i.e. standard metrics respond the same way to true and erroneous motion fields, OFC favours small motions, gradient constancy has poor convergence properties, etc.).
Scene
Norm
Cost function type
Cost function
Input Image
Response vs. Error
Response PDF
Scene
Norm
Cost function type
Cost function
Input Image
Response vs. Error
Response PDF
The webpage uses HTML5 and javascript. It was tested on all modern browsers (Chrome, Firefox, Safari and Internet Explorer 11). If there are issues, please check that javascript is enabled.
