CNN-ASSISTED COVERINGS IN THE SPACE OF TILTS: BEST AFFINE INVARIANT PERFORMANCES WITH THE SPEED OF CNNS

The classic approach to image matching consists in the detection, description and matching of keypoints. In description, the local information surrounding the keypoint is encoded. This locality encourages affine invariant methods. Indeed, smooth transformations (e.g. viewpoint deformations) are well approximated by affine maps. Despite numerous efforts, affine invariant descriptors remained elusive, ultimately resulting in the compromise of using viewpoint simulations to attain the desired invariance. Recent CNN-based methods seem to provide a way to learn affine invariance. Still, as a first contribution, we show that current CNN-based methods are still far from the state-of-the-art performance provided by IMAS (Image Matching by Affine Simulation) methods. Confirming that there is still room for improvement for learned methods. Second, we show that recent advances in affine patch normalization can be used to create adaptive IMAS methods that select their simulations depending on query and target images. The proposed methods are shown to attain a good compromise: on the one hand, they perform at the same level of the state-of-the-art IMAS methods but faster; and on the other hand, they perform significantly better than non-simulating methods, including recent ones.