Study of Convolutional Neural Networks for Global Parametric Motion Estimation on Log-Polar Imagery

The problem of motion estimation from images has been widely studied in the past. Although many mature solutions exist, there are still open issues and challenges to be addressed. For instance, in spite of the well-known performance of convolutional neural networks (CNNs) in many computer vision problems, only very recent work has started to explore CNNs to learning to estimate motion, as an alternative to manually-designed algorithms. These few initial efforts, however, have focused on conventional Cartesian images, while other imaging models have not been studied. This work explores the yet unknown role of CNNs in estimating global parametric motion in log-polar images. Despite its favourable properties, estimating some motion components in this model has proven particularly challenging with past approaches. It is therefore highly important to understand how CNNs behave when their input are log-polar images, since they involve a complex mapping in the motion model, a polar image geometry, and space-variant resolution. To this end, a CNN is considered in this work for regressing the motion parameters. Experiments on existing image datasets using synthetic image deformations reveal that, interestingly, standard CNNs can successfully learn to estimate global parametric motion on log-polar images with accuracies comparable to or better than with Cartesian images.