keypoint detection, neural networks, deep learning, panoramic images, SIFT

Pixel-precision level localization of keypoints is an essential step for stitching panoramic images as these keypoints are matching, and their locations are used for computing stitching transformation. We recall the main standard computer vision techniques for keypoint localization and focus on the precise localization. Based on the SIFT technique, we design a neural network architecture containing an encoder, a latent representation handler, and a decoder. In contrast to domain-agnostic neural network architectures, the developed encoder reflects the scale-space construction as well as the difference of Gaussians estimation used in SIFT. In the benchmark, we show that our architecture has a higher number of keypoints localized with pixel precision considering flips, intensity changes, and blurrings than other standard and neural network-based approaches.