Image Descriptor Learning for Matching Historical Aerial Images with Present-Day Satellite Images

Status: available
Supervisor: Sebastian Zambanini

Problem Statement

Learning local image descriptors by means of deep convolutional neural nets [1,2] has recently shown to produce stronger features than traditional hand-crafted ones such as SIFT [3]. However, these nets have been trained and evaluated on general scenarios of (wide-basline) object matching. For the DeVisOr project, matching historical aerial images with modern satellite images is of interest. In this scenario, different feature transformations have be learned due to the substantially different kinds of variations in image content.

Goal

The goal of this work is to apply modern feature learning techniques to the problem of matching historical aerial images with present-day satellite images. The work also includes the preparation of training data, for which already registered image pairs are available. Access to powerful deep learning servers will be granted.

Workflow

Literature Review – getting to know the algorithms
Data Preparation
Implementation
Evaluation
Written Report/Thesis and final presentation

Literature

[1] Simo-Serra, Edgar, et al. “Discriminative learning of deep convolutional feature point descriptors.” Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2015, pp. 118-126.

[2] Vijay Kumar B G, Gustavo Carneiro, Ian Reid, “Learning Local Image Descriptors With Deep Siamese and Triplet Convolutional Networks by Minimising Global Loss Functions” IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 5385-5394.

[3] Lowe, David G. “Distinctive image features from scale-invariant keypoints.” International Journal of Computer Vision, 60(2):91-110, 2004.