Feature Description and Detection with Deep Learning

Abstract

This thesis describes work with convolutional neural networks with the aim to detect and describe keypoints in images from a non-standard dataset, which is more complex than the datasets typically used for this task. Two approaches are explored -- the first one relies on a network covering the typical feature extraction pipeline, which consists of keypoint detection, orientation estimation and feature description, whereas the second approach uses an object detector network to detect and label specified keypoints.