In this paper, we describe our approach as part of the MediaEval 2014 Placing Task evaluation. We rst identify tags that are most indicative of geographic location by calculating the spatial-aware weighting for all tags in the training set. These weighting are applied to a language model-based retrieval framework. To address the geo-tagging problem, we nd the most similar training item and propagate its location to the test item. Base on last year's experience, we further improve the accuracy by utilizing the geo-location correlation of images/videos uploaded by the same user.
The MediaEval 2014 Placing Task requires participants to
assign geographical coordinates (latitude and longitude) to
Flickr images or videos (we denote them by Flickr items for
description simplicity), we refer to [
A total of 5;025;000 geo-referenced Flickr items are provided as training data. For language model-based approach, we treat each Flickr item's tags as a document. Other surrounded texts, such as title or description, are not used in our approach. We carried out two preliminary lter steps on this training set. First, items without tags were removed. Second, we converted all tags to lowercase and special characters were removed. Finally, this resulted in a pre-processed training set with 4;148;564 items. Unless speci ed, this pre-processed training set is used in the following experiments. 2.2
We use a Ripley's K statistic based tag selection method
[
Pp2Qt (jfqjq 2 Qt; q 6= p; d(p; q) gj) w
; Nt2 (1) where d( ) is the distance function. The weighting s(t) is similar to \tf-idf": the rst part log(Nt) will prefer tag with large frequency; the second part will downgrade the s(t) if tag t spreads all over the world and vise verse. Speci cally, when w = 1, if all the images with tag t cluster in a small region (controlled by ), the second part will near to 1, otherwise, near to 0. In practice, Qt doesn't need to contain all the items with tag t. For example, if there are more than 1 million Flickr items have t, we can only sample 5000 or so of them, which will be su cient enough to calculate the weighting. For each tag in the training set, we calculate its spatial-aware weighting by equation (1). 2.3
We use the framework proposed by [
To address the data sparsity issue of training data, [
Flickr users can organize their images and videos by assigning them to di erent collections (or albums). Intuitively, items within the same collection would be highly geo-correlated. Take Figure 1 as an example, a user shared his images during a trip to Brisbane, and organized them into a collection named Brisbane Trip 2014. As we can see, not every images in this collection is well tagged because user only tagged the images he loved or interested in, and leaving others un-tagged or poorly tagged. Moreover, it's di cult for us to predict their location by the image itself because none of them contain particular landmark or landscape. Images/videos with completely di erent tags or visual content could be considered as taken in the same location or nearby, if they were within the same user collection. For tag-based geo-tagging approaches, a poorly tagged query item will result in a bad estimation. However, if this item belongs to a user collection which contains one or more images/videos with well estimated location (usually well tagged or contain landmark), then we can use the centroid location of this collection as the estimation for the poorly estimated one.
In this paper we adopted similar strategy as last year we
did to nd test items within the same collection, please refer
[
There are ve di erent test sets and we chose test 5 whose
size is 510;000. Following [
We submitted two runs (run 1 and run 3) and the results of our experiments are shown in table 1. Run 2 is omitted which requires only visual and audio cues can be used. Baseline approach used the same retrieval model as run 1, but the spatial-aware tag weighting were not applied. Both baseline approach and run 1 assigned a default location, e.g., New York City (40:7127; 74:0059) in our case, for test items that without tag, whereas run 3 utilized the collection geocorrelation as discussed in section 2.4. As we can see, both spatial-aware tag weight and collection geo-correlation can help improve the geo-tagging accuracy. In this paper, we have set xed values for w, and and avoided tailoring these values to the problem. However, we believe there is potential for improvement in the results through the optimal selection of these parameters for the particular data. 4.