This paper addresses the challenge of dense pixel correspondence estimationbetween two images. This problem is closely related to optical flow estimationtask where ConvNets (CNNs) have recently achieved significant progress. Whileoptical flow methods produce very accurate results for the small pixeltranslation and limited appearance variation scenarios, they hardly deal withthe strong geometric transformations that we consider in this work. In thispaper, we propose a coarse-to-fine CNN-based framework that can leverage theadvantages of optical flow approaches and extend them to the case of largetransformations providing dense and subpixel accurate estimates. It is trainedon synthetic transformations and demonstrates very good performance to unseen,realistic, data. Further, we apply our method to the problem of relative camerapose estimation and demonstrate that the model outperforms existing denseapproaches.