Section: New Results

Robust motion model selection

Participants : Patrick Bouthemy, Bertha Mayela Toledo Acosta.

Parametric motion models are commonly used in image sequence analysis for different tasks. A robust estimation framework is usually required to reliably compute the motion model. However, choosing the most appropriate model in that estimation context is still an open issue. Indeed, penalizing the model complexity while maximizing the size of the inlier set may be contradictory. In this study, we proposed a robust motion model selection method which relies on the Fisher statistic. We also derived an interpretation of it as a robust $C_P$-Mallows criterion. The resulting criterion is straightforward to compute and explicitly involves the aforementioned trade-off between maximizing the size of the inlier set and minimizing the complexity (i.e., the number of parameters) of the selected motion model. We have conducted a comparative experimental evaluation on synthetic and real image sequences demonstrating that our criterion outperforms the RBIC criterion.

Collaborator: Bernard Delyon (IRMAR Rennes).