Today's geo-location estimation approaches are able to infer the location of a target image using its visual content alone. These approaches typically exploit visual matching techniques, applied to a large collection of background images with known geo-locations. Users who are unaware that visual analysis and retrieval approaches can compromise their geo-privacy, unwittingly open themselves to risks of crime or other unintended consequences. This paper lays the groundwork for a new approach to geo-privacy of social images: Instead of requiring a change of user behavior, we start by investigating users' existing photo-sharing practices. We carry out a series of experiments using a large collection of social images (8.5M) to systematically analyze how photo editing practices impact the performance of geo-location estimation. We find that standard image enhancements, including filters and cropping, already serve as natural geo-privacy protectors. In our experiments, up to 19% of images whose location would otherwise be automatically predictable were unlocalizeable after enhancement. We conclude that it would be wrong to assume that geo-visual privacy is a lost cause in today's world of rapidly maturing machine learning. Instead, protecting users against the unwanted effects of pixel-based inference is a viable research field. A starting point is understanding the geo-privacy bonus of already established user behavior.

This paper presents a novel two-phase method for audio representation, discriminative and compact audio representation (DCAR), and evaluates its performance at detecting events and scenes in consumer-produced videos. In the first phase of DCAR, each audio track is modeled using a Gaussian mixture model (GMM) that includes several components to capture the variability within that track. The second phase takes into account both global structure and local structure. In this phase, the components are rendered more discriminative and compact by formulating an optimization problem on a Grassmannian manifold. The learned components can effectively represent the structure of audio. Our experiments used the YLI-MED and DCASE Acoustic Scenes datasets. The results show that variants on the proposed DCAR representation consistently outperform four popular audio representations (mv-vector, i-vector, GMM, and HEM-GMM). The advantage is significant for both easier and harder discrimination tasks; we discuss how these performance differences across tasks follow from how each type of model leverages (or does not leverage) the intrinsic structure of the data.

This paper presents a novel two-phase method for audio representation: Discriminative and Compact Audio Representation (DCAR). In the first phase, each audio track is modeled using a Gaussian mixture model (GMM) that includes several components to capture the variability within that track. The second phase takes into account both global structure and local structure. In this phase, the components are rendered more discriminative and compact by formulating an optimization problem on Grassmannian manifolds, which we found represents the structure of audio effectively. Experimental results on the YLI-MED dataset show that the proposed DCAR representation consistently outperforms state-of-the-art audio representations: i-vector, mv-vector, and GMM.

We present Evento 360 (URL: http://evento360.info), an online interactive social event browser, which allows the user to explore events detected within a web-scale multimedia corpus. The system addresses five key aspects of social multimedia event detection and summarization: multimodality, scale, diversity of representations, noise of multimedia items, and missing metadata. The detection algorithm uses unsupervised clustering approach that exploits temporal, spatial and textual metadata. For each detected event cluster, to choose the best subset of photos that meet both relevance and diversity criteria, the system uses hierarchical clustering that exploits both visual and audio information. Evento 360's user interface provides a search feature that is not limited to a certain set of events, but rather can handle an arbitrary event query. It allows the user to retrieve and explore relevant events. The system scales well and is effective in producing high-quality summaries of the detected events.

Crowdsourcing is currently used for a range of applications, either by exploiting unsolicited user-generated content, such as spontaneously annotated images, or by utilizing explicit crowdsourcing platforms such as Amazon Mechanical Turk to mass-outsource artificial-intelligence-type jobs. However, crowdsourcing is most often seen as the best option for tasks that do not require more of people than their uneducated intuition as a human being. This article describes our methods for identifying workers for crowdsourced tasks that are difficult for both machines and humans. It discusses the challenges we encountered in qualifying annotators and the steps we took to select the individuals most likely to do well at these tasks.

This work presents improved audio-based user-verification analysis and results on Flickr videos, using a subset of the MediaEval 2011 [1] data set. User-verification is a new task, where the goal is to determine if two pieces of media are uploaded by the same user. Our best results, with a 19.7% Equal Error Rate, and a 53.9% Miss Rate at 1% False Positive, are obtained using an i-vector [2] system. A frequency-matching system that requires 96% less computation time than the other systems is also explored, and may be better suited for processing large datasets from Flickr and other social networks. The results have significant privacy implications as they present a framework for exploiting users' tendencies to assume that different accounts remain as separate realms.

Over the recent years, the problem of video location estimation (i.e., estimating the longitude/latitude coordinates of a video without GPS information) has been approached with diverse methods and ideas in the research community and significant improvements have been made. So far, however, systems have only been compared against each other and no systematic study on human performance has been conducted. Based on a human-subject study with 11,900 experiments, this article presents a human baseline for location estimation for different combinations of modalities (audio, audio/video, audio/video/text). Furthermore, this article compares state-of-the-art location estimation systems with the human baseline. Although the overall performance of humans' multimodal video location estimation is better than current machine learning approaches, the difference is quite small: For 41% of the test set, the machine's accuracy was superior to the humans. We present case studies and discuss why machines did better for some videos and not for others. Our analysis suggests new directions and priorities for future work on the improvement of location inference algorithms.

We have performed city-verification of videos based on the videos' audio and metadata, using videos from the MediaEval Placing Task's video set, which contain consumer-produced videos “from-the-wild”. 18 cities were used as targets, for which acoustic and language models were trained, and against which test videos were scored. We have obtained the first known results for the city verification task, with an EER minimum of 21.8%, suggesting that ~80% of test videos, when tested against a correct target city, were identified as belonging to that city. This result is well above-chance, even as the videos contained very few city-specific audio and metadata features. We have also demonstrated the complementarity of audio and metadata for this task.

This article describes a novel approach to the problem of associating geo-locations to consumer-produced multimedia data such as videos and photos that are publicly available on social networking websites such as Flickr. We specifically focus on the case where the available training data is sparse both in absolute numbers as well as geographic coverage when compared to the number of untagged query data. We develop a novel graphical model based framework for the problem of interest and pose the problem of geotagging as one of inference over this graph. The novelty of our algorithm lies in the fact that we jointly estimate the geo-locations of all the query videos, which helps obtain performance improvements over existing algorithms in the literature that process each query video independently. Our system enables the query videos to act as "virtual" training data that effectively bootstrap the geo-tagging process. The quality of the database improves with each additional query video in the system. Further, our modeling provides a generic theoretical framework that can be used to incorporate any other available textual, visual or audio features. We evaluate our algorithm on the MediaEval 2011 Placing Task data set and show that for fixed training data the system performance improves with an increasing number of unlabeled test data. The performance gains are shown to be over 10% as compared to existing algorithms in the literature.

This article presents an attempt to link the uploaders of videos based on the audio track of the videos. Using a subset of the MediaEval Placing Task's Flickr video set, which is labeled with the uploader's name, we conducted an experiment with a similar setup as a typical NIST speaker recognition evaluation run. Based on the assumption that the audio might be matched in various ways (speaker, channel, environmental noise, etc.), we trained one of ICSI's simplified speaker recognition systems on the audio tracks of the Flickr videos. Note that since the selection of videos is essentially random, the audio track can contain any sounds. We obtain an equal error rate of 36.7% on 312 videos with 11,550 trials. The result has implications for audio research, security applications, and raises privacy concerns.

The following article describes our demo of an approach to determine the geo-coordinates of the recording place of Flickr videos based on both textual metadata and visual cues. The underlying system has been tested on the MediaEval 2010 Placing Task evaluation data, which consists of 5091 unfiltered test videos is able to classify 14% of the videos to within an accuracy of 10m.