A framework for dense optical flow from multiple sparse hypotheses

Optical flow forms an important initial processing stage for many machine vision tasks. A framework is presented for the recovery of dense optical flows from image sequences containing large motions. Sparse feature correspondences are used to assign multiple optical flow hypotheses to each image pixel which are then independently refined to produce a further set of refined hypotheses. One final flow is selected for each pixel from these refined flows by seeking to minimize the local matching error. Dense optical flows from image sequences with small motions are successfully recovered. In image sequences with very large motions, a clear increase in optical flow accuracy is observed when compared to a hierarchical approach to optical flow estimation.