Center for Research in Comptuer Vision
Center for Research in Comptuer Vision

GPS Tag Refinement Using Random Walks and Adaptive Damping (CVPR 2014 - Download PDF,)

Why GPS Tag Refinement?

Due to the emergence of mobile devices with various internal positioning methods, the majority of images taken nowadays can be geo-tagged at the time of collection. However, different tagging systems, e.g. GPS, WiFi positioning system (WPS), Cell positioning system, manual tagging, etc. have a broad range of reliability and accuracy which altogether translate into inaccuracies in the geo-tags of usershared images. In this paper, we propose a method for GPS-tag refinement . That is, given a dataset of GPS-tagged images with an unknown subset which includes inaccuracies in the tags, we discover the contaminated subset and adjust its GPS-tags to the correct locations. We achieve this goal utilizing the rest of the images in the dataset (i.e. self-refinement) without using any other source of imagery or data.



Our Approach

We refine the GPS-tags utilizing the rest of the images in the dataset (i.e. self-refinement) without using any other source of imagery or data. We accomplish this task by generating a large number of estimations for the location of a particular image in the dataset based on the rest of the images therein. Then, we use a robust method based on Random Walks which identifies the reliable estimations based on their pairwise consistency and use them for computing the refined GPS-tag.

Image Matching and Location Estimation

We match the query image against the rest of the images in the dataset and retrieve matches using bag of words method with a vocabulary size of 50k along with the tf-idf voting scheme for matching image triplets using the query image and each possible pair of the retrieved matches are formed. We estimate the relative location of the query image with respect to the two matched images by finding the trifocal tensor and performing structure from motion


The estimation of the GPS-location of the image which a triplet yields is accurate only if both of the parent reference images have an accurate GPS-tag. Thus, since we assume an unknown subset of the GPS-tags of the images in the dataset are inaccurate, a considerable number of the estimations are incorrect. However, the correct estimations are expected to show a high consistency with each other whereas the incorrect one are more less randomly distributed. Therefore, we use random walk for discovering the reliable subset. We also propose an adaptive damping factor for random walks so that it conforms with the amount of the noise in the data.


User-shared images typically show a severely non-uniform geo-distribution; this characteristic can potentially result in a reduction in the accuracy of tag-refinement. When performing image matching between the query and the dataset, more images from the popular spot are likely to be retrieved as more images from there exist in the dataset. Consequently, there will be more triplet estimations arisen from that spot and the final estimation of the random walk will be leaning towards the location suggested by the images of the popular spot. This deviation is merely due to the reason that more images from that spot exist in the database, and not the query-related operations; therefore, it is undesirable as it acts as a bias. In order to reduce the impact of this phenomena, we incorporate the density of the dataset in our random walk formulation.


The following figure illustrates the results of tag refinement for different amount of noise in the initial tags:

The following shows effect of the adaptive damping on the output error:


The following figures show the robustness evaluation of our method and its comparison to baseline methods for mode seeking:


Random Walks with Adaptive Damping ,

Related Publications

Amir Roshan Zamir, Shervin Ardeshir and Mubarak Shah, "Robust Refinement of GPS-Tags using Random Walks with an Adaptive Damping Factor", in Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), June 2014 [PDF | Project Page]

Take a look at a few of our other papers and projects in the area of geo-spatial analysis of images and videos:

Shervin Ardeshir, Amir Roshan Zamir, Alejandro Torroella, and Mubarak Shah, "GIS-Assisted Object Detection and Geospatial Localization", in Proceedings of IEEE European Conference on Computer Vision (ECCV), September 2014 [PDF | Project Page]

Amir Roshan Zamir, Afshin Dehghan and Mubarak Shah, "Visual Business Recognition - A Multimodal Approach", In Proceeding of ACM International Conf. on Multimedia (ACM MM), 2013 [PDF | BibTeX | Supplemental Mat.]

Amir Roshan Zamir and Mubarak Shah, "Image Geo-localization Based on Multiple Nearest Neighbor Feature Matching using Generalized Graphs", In IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2014 [PDF | Project Page]

Gonzalo Vaca, Amir Roshan Zamir and Mubarak Shah, "City Scale Geo-spatial Trajectory Estimation of a Moving Camera", in Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), June 2012 [PDF | BibTeX| Project Page]

Amir Roshan Zamir and Mubarak Shah, "Accurate Image Localization Based on Google Maps Street View", In Proceedings of European Conference on Computer Vision (ECCV), 2010 [PDF | BibTeX | Project Page]

Back to Geospatial Localization / Registration Projects