On Dynamic Scene Geometry for View-Invariant Action Matching

Abstract

Variation in viewpoints poses significant challenges to action recognition. One popular way of encoding view-invariant action representation is based on the exploitation of epipolar geometry between different views of the same action. Majority of representative work considers detection of landmark points and their tracking by assuming that motion trajectories for all landmark points on human body are available throughout the course of an action. Unfortunately, due to occlusion and noise, detection and tracking of these landmarks is not always robust. To facilitate it, some of the work assumes that such trajectories are manually marked which is a clear drawback and lacks automation introduced by computer vision. In this paper, we address this problem by proposing view invariant action matching score based on epipolar geometry between actor silhouettes, without tracking and explicit point correspondences. In addition, we explore multi-body epipolar constraint which facilitates to work on original action volumes without any pre-processing. We show that multi-body fundamental matrix captures the geometry of dynamic action scenes and helps devising an action matching score across different views without any prior segmentation of actors. Extensive experimentation on challenging view invariant action datasets shows that our approach not only removes long standing assumptions but also achieves significant improvement in recognition accuracy and retrieval.