Forward-Backward Consistent Trajectories in Video Sequences

This paper is concerned with accuracy estimation of point trajectories in video sequences without ground truth information. This is an essential problem for many computer vision applications which consider the point trajectories as an abstraction layer for further processing. We exploit the forward-backward consistency principle and design a novel measure called k-step error: the same tracking algorithm is applied on a reversed sequence of images and creates a validation trajectory which is compared to the trajectory in question. The contribution of the paper is in the quantitative evaluation of the proposed measure against commonly used SSD error and demonstration of its applicability on two problems: (i) selection of reliable feature points and (ii) real-time object tracking. A significant improvement in terms of tracker robustness in comparison with state-ofthe-art is demonstrated on benchmark video sequences.

• Type: Conference
• Year: 2010
• Where: ICPR
• Authors: Zdenek Kalal, Krystian Mikolajczyk, Jiri Matas
• Keywords: Computer Vision, Forward-backward Consistency Principle, ICPR 2010, Point Trajectories, Video Sequences