Image sharing | EPFL Graph Search

Image sharing, or photo sharing, is the publishing or transfer of digital photos online. Image sharing websites offer services such as uploading, hosting, managing and sharing of photos (publicly or privately). This function is provided through both websites and applications that facilitate the upload and display of images. The term can also be loosely applied to the use of online photo galleries that are set up and managed by individual users, including photoblogs. Sharing means that other users can view but not necessarily download images, and users can select different copyright options for their images. While photoblogs tend only to display a chronological view of user-selected medium-sized photos, most photo sharing sites provide multiple views (such as thumbnails and slideshows), the ability to classify photos into albums, and add annotations (such as captions or tags). Desktop photo management applications may include their own photo-sharing features or integration with sites

Image Localization using Google Street View

Raffaele Lauro

Digital images, taken for example with a smartphone, are usually geo-tagged with location and viewing direction information. This information is not always accurate due to, for example, GPS inaccuracies in large cities. The aim of this project is to develop an algorithm that takes an image and estimates the location and viewing direction by key-point matching with Google street-view data. As a first step, the original image will be geo-tagged allowing an easy pruning of the street view data. The project could then progress to localizing a non-geo-tagged image using some additional location information, such as a street name or point of interest. This type of algorithm could have applications localizing the large number of images on the Internet with unreliable or no geo-tagging information. In addition, one can envision such a system being employed in indoor environments where GPS currently fails. The student will be expected to deliver software to take an image with imprecise geo-tagging and match it to Google street images with a similar location. The algorithm should return the location and orientation of the best matching street-view image. There are two possible extensions to this main requirement: 1. Use matches in multiple street-view images to estimate a more accurate location and orientation, between the street-view images. 2. Improve the recognition stage so that the algorithm can deal with less accurate geo-tagging.

2015

Privacy-Friendly Photo Sharing and Relevant Applications Beyond

Lin Yuan

Popularization of online photo sharing brings people great convenience, but has also raised concerns for privacy. Researchers proposed various approaches to enable image privacy, most of which focus on encrypting or distorting image visual content. In this thesis, we investigate novel solutions to protect image privacy with a particular emphasis on online photo sharing. To this end, we propose not only algorithms to protect visual privacy in image content but also design of architectures for privacy-preserving photo sharing. Beyond privacy, we also explore additional impacts and potentials of employing daily images in other three relevant applications. First, we propose and study two image encoding algorithms to protect visual content in image, within a Secure JPEG framework. The first method scrambles a JPEG image by randomly changing the signs of its DCT coefficients based on a secret key. The second method, named JPEG Transmorphing, allows one to protect arbitrary image regions with any obfuscation, while secretly preserving the original image regions in application segments of the obfuscated JPEG image. Performance evaluations reveal a good degree of storage overhead and privacy protection capability for both methods, and particularly a good level of pleasantness for JPEG Transmorphing, if proper manipulations are applied. Second, we investigate the design of two architectures for privacy-preserving photo sharing. The first architecture, named ProShare, is built on a public key infrastructure (PKI) integrated with a ciphertext-policy attribute-based encryption (CP-ABE), to enable the secure and efficient access to user-posted photos protected by Secure JPEG. The second architecture is named ProShare S, in which a photo sharing service provider helps users make photo sharing decisions automatically based on their past decisions using machine learning. The photo sharing service analyzes not only the content of a user's photo, but also context information about the image capture and a prospective requester, and finally makes decision whether or not to share a particular photo to the requester, and if yes, at which granularity. A user study along with extensive evaluations were performed to validate the proposed architecture. In the end, we research into three relevant topics in regard to daily photos captured or shared by people, but beyond their privacy implications. In the first study, inspired by JPEG Transmorphing, we propose an animated JPEG file format, named aJPEG. aJPEG preserves its animation frames as application markers in a JPEG image and provides smaller file size and better image quality than conventional GIF. In the second study, we attempt to understand the impact of popular image manipulations applied in online photo sharing on evoked emotions of observers. The study reveals that image manipulations indeed influence people's emotion, but such impact also depends on the image content. In the last study, we employ a deep convolutional neural network (CNN), the GoogLeNet model, to perform automatic food image detection and categorization. The promising results obtained provide meaningful insights in design of automatic dietary assessment system based on multimedia techniques, e.g. image analysis.

EPFL2017

Phrase-based Image Captioning

Ronan Collobert, Rémi Philippe Lebret, Pedro Henrique Oliveira Pinheiro

Generating a novel textual description of an image is an interesting problem that connects computer vision and natural language processing. In this paper, we present a simple model that is able to generate descriptive sentences given a sample image. This model has a strong focus on the syntax of the descriptions. We train a purely bilinear model that learns a metric between an image representation (generated from a previously trained Convolutional Neural Network) and phrases that are used to described them. The system is then able to infer phrases from a given image sample. Based on caption syntax statistics, we propose a simple language model that can produce relevant descriptions for a given test image using the phrases inferred. Our approach, which is considerably simpler than state-of-the-art models, achieves comparable results in two popular datasets for the task: Flickr30k and the recently proposed Microsoft COCO.

2015

Privacy-Friendly Photo Sharing and Relevant Applications Beyond

Lin Yuan

EPFL2017