Revaud Jerome Weinzaepfel Philippe Harchaoui Zaid Schmid Cordelia

Abstract

We introduce a novel matching algorithm, called DeepMatching, to computedense correspondences between images. DeepMatching relies on a hierarchical,multi-layer, correlational architecture designed for matching images and wasinspired by deep convolutional approaches. The proposed matching algorithm canhandle non-rigid deformations and repetitive textures and efficientlydetermines dense correspondences in the presence of significant changes betweenimages. We evaluate the performance of DeepMatching, in comparison withstate-of-the-art matching algorithms, on the Mikolajczyk (Mikolajczyk et al2005), the MPI-Sintel (Butler et al 2012) and the Kitti (Geiger et al 2013)datasets. DeepMatching outperforms the state-of-the-art algorithms and showsexcellent results in particular for repetitive textures.We also propose amethod for estimating optical flow, called DeepFlow, by integratingDeepMatching in the large displacement optical flow (LDOF) approach of Brox andMalik (2011). Compared to existing matching algorithms, additional robustnessto large displacements and complex motion is obtained thanks to our matchingapproach. DeepFlow obtains competitive performance on public benchmarks foroptical flow estimation.