The objective of this work is to develop an unsupervised deep learning-based method for shape retrieval. Given a query shape, the method should retrieve all shapes from the same class in a data set. The method should use specific scale- and rotation-invariant keypoints detected on discrete contours. Each keypoint is described by a feature vector and has the following geometrical information: position, scale, and orientation. The method should learn the underlying patterns and relationships among the keypoints in a general manner. For this, the following steps are planned: creating training data in the form of graphs, which represent the relationships among keypoints using their feature vectors and geometrical information to model the relative spatial arrangements of keypoints; developing and training an appropriate graph deep learning architecture; and calculating shape similarity. The developed method will be evaluated using the MPEG-7 data set, with the aim of achieving a high bull's eye score.