BENEFITS OF TEMPORAL OVERSAMPLING IN OPTICAL FLOW ESTIMATION (WA-P1)

Author(s) : SukHwan Lim (Hewlett-Packard Laboratories, USA) John Apostolopoulos (Hewlett-Packard Laboratories, USA) Abbas El Gamal (Stanford University, USA)

Abstract : Recently it has been shown that temporal oversampling can improve the accuracy of optical flow estimation (OFE). This paper identifies the factors that contribute to the improvement and attempts to quantify their contributions. Experiments performed with real video sequences illustrate that motion aliasing leads to inaccurate optical flow at the standard frame rate, which can be overcome through the use of higher frame rates. To quantify the effect of motion aliasing, we use synthetic sequences and determine the minimum frame rate that achieves good OFE accuracy as a function of motion and spatial frequency. In addition, if 2-tap temporal gradient filters are used for gradient-based OFE methods, we show that the capture frame rate should be 50% higher than that needed by ideal gradient estimators. The use of a sequence produced by a natural image undergoing synthetic (known) motion demonstrates that as the oversampling is increased, the motion-aliased energy is reduced and the OFE accuracy is increased. Therefore, temporal oversampling is a valuable mechanism for improving OFE performance, where the minimum frame rate is largely determined by the rate necessary to prevent motion aliasing in the sequence and accurately capture the high temporal frequencies.

Menu