a) Can you see the swimming people? (b) Where is the ball?Figure 1: Using a relational non-local module directly on the feature maps in a coarse-to-fine manner enables the detection of small objects, based on (i) repeating instances of the same class and (ii) the existence of larger related objects, allowing us to: (a) pay attention to the tiny swimmers in the sea and (b) locate the ball. Cyan -NL, Redours, ENL. Best viewed in color.

In this paper, we introduce a novel deep-learning method to align cross-spectral images. Our approach relies on a learned descriptor which is invariant to different spectra. Multi-modal images of the same scene capture different signals and therefore their registration is challenging and it is not solved by classic approaches. To that end, we developed a feature-based approach that solves the visible (VIS) to Near-Infra-Red (NIR) registration problem. Our algorithm detects corners by Harris [9] and matches them by a patchmetric learned on top of a network trained on CIFAR-10 [12] dataset. As our experiments demonstrate we achieve a high-quality alignment of cross-spectral images with a subpixel accuracy. Comparing to other existing methods, our approach is more accurate in the task of VIS to NIR registration.

No abstract

No abstract

No abstract