MOLINA-CABELLO ET AL.: HOMOGRAPHY BY RANDOM COLOR TRANSFORMATIONS 1 Homography estimation with deep convolutional neural networks by random color transformations Miguel A. Molina-Cabello 1 miguelangel@lcc.uma.es David A. Elizondo 2 elizondo@dmu.ac.uk Rafael Marcos Luque-Baena 1 rmluque@lcc.uma.es Ezequiel López-Rubio 1 ezeqlr@lcc.uma.es 1 Department of Computer Languages and Computer Science University of Málaga Málaga, Spain 2 Department of Computer Technology De Montfort University Leicester, United Kingdom Abstract Most classic approaches to homography estimation are based on the filtering of out- liers by means of the RANSAC method. New proposals include deep convolutional neu- ral networks. Here a new method for homography estimation is presented, which supplies a deep neural homography estimator with color perturbated versions of the original im- age pair. The obtained outputs are combined in order to obtain a more robust estimation of the underlying homography. Experimental results are shown, which demonstrate the adequate performance of our approach, both in quantitative and qualitative terms. 1 Introduction One of the fundamental low level tasks to be accomplished in computer vision is that of homography estimation between two images. This task consists in finding a non-singular, linear homography transformation between the points in both images [5, 6]. This is required for many higher level tasks of computer vision such as line matching [15], mosaicing [3], motion detection [18], camera motion estimation [19], tracking from multiple views [7, 8, 9, 10] and action recognition [17]. Classic approaches to homography estimation are mostly based on the Random Sampling and Consensus (RANSAC) technique, where a filtering of pairs of key points coming from both images is carried out, in order to remove erroneous point pairs [1, 2, 14]. Once the bad point pairs have been identified, the estimation of the homography can proceed more accurately. Outlier rejection is therefore the main strategy in which classic approaches are founded [13, 16]. This state of things has changed substantially with the advent of deep learning convolu- tional neural networks. Deep homography estimation networks have been proposed, which c  2019. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms.