Video summarization and video captioning are considered two separate tasks in existing studies. For longer videos, automatically identifying the important parts of video content and annotating them with captions will enable a richer and more concise condensation of the video. We propose a general neural network architecture that jointly considers two supervisory signals (i.e., an image-based video summary and text-based video captions) in the training phase and generates both a video summary and corresponding captions for a given video in the test phase. Our main idea is that the summary signals can help a video captioning model learn to focus on important frames. On the other hand, caption signals can help a video summarization model to learn better semantic representations. Jointly modeling both the video summarization and the video captioning tasks offers a novel end-to-end solution that generates a captioned video summary enabling users to index and navigate through the highlights in a video. Moreover, our experiments show the joint model can achieve better performance than state-of- the-art approaches in both individual tasks.

In this paper, we describe DocHandles, a novel system that allows users to link to specific document parts in their chat applications. As users type a message, they can invoke the tool by referring to a specific part of a document, e.g., “@fig1 needs revision”. By combining text parsing and document layout analysis, DocHandles can find and present all the figures “1” inside previously shared documents, allowing users to explicitly link to the relevant “document handle”. Documents become first-class citizens inside the conversation stream where users can seamlessly integrate documents in their text-centric messaging application.

It is increasingly possible to use cameras and sensors to detect and analyze human appearance for the purposes of personalizing user experiences. Such systems are already deployed in some public places to personalize advertisements and recommend items. However, since these technologies are not yet widespread, we do not have a good sense of the perceived benefits and drawbacks of public display systems that use face detection as an input for personalized recommendations. We conducted a user study with a system that inferred a user’s gender and age from a facial detection and analysis algorithm and used this to present recommendations in two scenarios (finding stores to visit in a mall and finding a pair of sunglasses to buy).  This work provides an initial step towards understanding user reactions to a new and emerging form of implicit recommendation based on physical appearance.

The availability of mobile access has shifted social media use. With that phenomenon, what users shared on social media and where they visited is naturally an excellent resource to learn their visiting behavior. Knowing visit behaviors would help market survey and customer relationship management, e.g., sending customers coupons of the businesses that they visit frequently. Most prior studies leverage meta-data e.g., check- in locations to profile visiting behavior but neglect important information from user-contributed content, e.g., images. This work addresses a novel use of image content for predicting the user visit behavior, i.e., the frequent and regular business venue categories that the content owner would visit. To collect training data, we propose a strategy to use geo-metadata associated with images for deriving the labels of an image owner’s visit behavior. Moreover, we model a user’s sequential images by using an end-to-end learning framework to reduce the optimization loss. That helps improve the prediction accuracy against the baseline as demonstrated in our experiments. The prediction is completely based on image content that is more available in social media than geo-metadata, and thus allows coverage in profiling a wider set of users.

Video conferencing is widely used to help deliver educational presentations, such as lectures or informational webinars, to a distributed audience. While individuals in a dyadic conversation may be able to use webcam streams to assess the engagement level of their interlocutor with some ease, as the size of the audience in a video conference setting increases, it becomes increasingly difficult to interpret how engaged the overall group may be. In this work, we use a mixed-methods approach to understand how presenters and attendees of online presentations use available cues to perceive and interpret audience behavior (such as how engaged the group is). Our results suggest that while webcams are seen as useful by presenters to increase audience visibility and encourage attention, audience members do not uniformly benefit from seeing others’ webcams; other interface cues such as chat may be more useful and informative engagement indicators for both parties. We conclude with design recommendations for future systems to improve what is sensed and presented.

In this paper, we propose a real-time classification scheme to cope with noisy Radio Signal Strength Indicator (RSSI) measurements utilized in indoor positioning systems. RSSI values are often converted to distances for position estimation. However due to multipathing and shadowing effects, finding a unique sensor model using both parametric and nonparametric methods is highly challenging. We learn decision regions using the Gaussian Processes classification to accept measurements that are consistent with the operating sensor model. The proposed approach can perform online, does not rely on a particular sensor model or parameters, and is robust to sensor failures. The experimental results achieved using hardware show that available positioning algorithms can benefit from incorporating the classifier into their measurement model as a meta-sensor modeling technique.

Observe at a person pointing out and describing something. Where is that person looking? Chances are good that this person also looks at what she is talking about and pointing at. Gaze is naturally coordinated with our speech and hand movements. By utilizing this tendency, we can create a natural interaction with computing devices and environments. In this chapter, we will first briefly discuss some basic properties of the gaze signal we can get from eye trackers, followed by a review of a multimodal system utilizing the gaze signal as one input modality. In Multimodal Gaze Interaction, data from eye trackers is used as an active input mode where for instance gaze is used as an alternative, or complimentary, pointing modality along with other input modalities. Using gaze as an active or explicit input method is challenging for several reasons. One of them being that eyes are primarily used for perceiving our environment, so knowing when a person selects an item with gaze versus just looking around is an issue. Researchers have tried to solve this by combining gaze with various input methods, such as manual pointing, speech, touch, etc. However, gaze information can also be used in interactive systems, for other purposes than explicit pointing since a user's gaze is a good indication of the user's attention. In passive gaze interaction, the gaze is not used as the primary input method, but as a supporting input method. In these kinds of systems, gaze is mainly used for inferring and reasoning about the user's cognitive state or activities in a way that can support the interaction. These kinds of multimodal systems often combine gaze with a multitude of input modalities. In this chapter we focus on interactive systems, exploring the design space for gaze-informed multimodal interaction spanning from gaze as active input mode to passive and if the usage scenario is stationary (at e.g. a desk) or mobile. There are a number of studies aimed at describing, detecting or modeling specific behaviors or cognitive states. We will touch on some of these works since they can guide us in how to build gaze-informed multimodal interaction.

Work breaks can play an important role in the mental and physical well-being of workers and contribute positively to productivity. In this paper we explore the use of activity-, physiological-, and indoor-location sensing to promote mobility during work-breaks. While the popularity of devices and applications to promote physical activity is growing, prior research highlights important constraints when designing for the workplace. With these constraints in mind, we developed BreakSense, a mobile application that uses a Bluetooth beacon infrastructure, a smartphone and a smartwatch to encourage mobility during breaks with a game-like design. We discuss constraints imposed by design for work and the workplace, and highlight challenges associated with the use of noisy sensors and methods to overcome them. We then describe a short deployment of BreakSense within our lab that examined bound vs. unbound augmented breaks and how they affect users’ sense of completion and readiness to work.

Users often use social media to share their interest in products. We propose to identify purchase stages from Twitter data following the AIDA model (Awareness, Interest, Desire, Action). In particular, we define a task of classifying the purchase stage of each tweet in a user's tweet sequence. We introduce RCRNN, a Ranking Convolutional Recurrent Neural Network which computes tweet representations using convolution over word embeddings and models a tweet sequence with gated recurrent units. Also, we consider various methods to cope with the imbalanced label distribution in our data and show that a ranking layer outperforms class weights.

We present Lift, a visible light-enabled finger tracking and object localization technique that allows users to perform freestyle multi-touch gestures on any object’s surface in an everyday environment. By projecting encoded visible patterns onto an object’s surface (e.g. paper, display, or table), and localizing the user’s fingers with light sensors, Lift offers users a richer interactive space than the device’s existing interfaces. Additionally, everyday objects can be augmented by attaching sensor units onto their surface to accept multi-touch gesture input. We also present two applications as a proof of concept. Finally, results from our experiments indicate that Lift can localize ten fingers simultaneously with accuracy of 0.9 mm and 1.8 mm on two axes respectively and an average refresh rate of 84 Hz with 16.7ms delay on WiFi and 12ms delay on serial, making gesture recognition on noninstrumented objects possible.

This is a summary of our participation in the TRECVID 2016 video hyperlinking task (LNK). We submitted four runs in total. A baseline system combined on established vectorspace text indexing and cosine similarity. Our other runs explored the use of distributed word representations in combination with fine-grained inter-segment text similarity measures.

Videos recorded with current mobile devices are increasingly geotagged at fine granularity and used in various location based applications and services. However, raw sensor data collected is often noisy, resulting in subsequent inaccurate geospatial analysis. In this study, we focus on the challenging correction of compass readings and present an automatic approach to reduce these metadata errors. Given the small geo-distance between consecutive video frames, image-based localization does not work due to the high ambiguity in the depth reconstruction of the scene. As an alternative, we collect geographic context from OpenStreetMap and estimate the absolute viewing direction by comparing the image scene to world projections obtained with different external camera parameters. To design a comprehensive model, we further incorporate smooth approximation and feature-based rotation estimation when formulating the error terms. Experimental results show that our proposed pyramid-based method outperforms its competitors and reduces orientation errors by an average of 58.8%. Hence, for downstream applications, improved results can be obtained with these more accurate geo-metadata. To illustrate, we present the performance gain in landmark retrieval and tag suggestion by utilizing the accuracy-enhanced geo-metadata.

Accurate map matching has been a fundamental but challenging problem that has drawn great research attention in recent years. It aims to reduce the uncertainty in a trajectory by matching the GPS points to the road network on a digital map. Most existing work has focused on estimating the likelihood of a candidate path based on the GPS observations, while neglecting to model the probability of a route choice from the perspective of drivers. Here we propose a novel feature-based map matching algorithm that estimates the cost of a candidate path based on both GPS observations and human factors. To take human factors into consideration is very important especially when dealing with low sampling rate data where most of the movement details are lost. Additionally, we simultaneously analyze a subsequence of coherent GPS points by utilizing a new segment-based probabilistic map matching strategy, which is less susceptible to the noisiness of the positioning data. We have evaluated the proposed approach on a public large-scale GPS dataset, which consists of 100 trajectories distributed all over the world. The experimental results show that our method is robust to sparse data with large sampling intervals (e.g., 60 s to 300 s) and challenging track features (e.g., u-turns and loops). Compared with two state-of-the-art map matching algorithms, our method substantially reduces the route mismatch error by 6.4% to 32.3% and obtains the best map matching results in all the different combinations of sampling rates and challenging features.

Improvements in sensor and wireless network enable accurate, automated, instant determination and dissemination of a user's or objects position. The new enabler of location-based services (LBSs) apart from the current ubiquitous networking infrastructure is the enrichment of the different systems with semantics information, such as time, location, individual capability, preference and more. Such semantically enriched system-modeling aims at developing applications with enhanced functionality and advanced reasoning capabilities. These systems are able to deliver more personalized services to users by domain knowledge with advanced reasoning mechanisms, and provide solutions to problems that were otherwise infeasible. This approach also takes user's preference and place property into consideration that can be utilized to achieve a comprehensive range of personalized services, such as advertising, recommendations, or polling. This paper provides an overview of indoor localization technologies, popular models for extracting semantics from location data, approaches for associating semantic information and location data, and applications that may be enabled with location semantics. To make the presentation easy to understand, we will use a museum scenario to explain pros and cons of different technologies and models. More specifically, we will first explore users' needs in a museum scenario. Based on these needs, we will then discuss advantages and disadvantages of using different localization technologies to meet these needs. From these discussions, we can highlight gaps between real application requirements and existing technologies, and point out promising localization research directions. By identifying gaps between various models and real application requirements, we can draw a road map for future location semantics research.

We propose a robust pointing detection with virtual shadow representation for interacting with a public display. Using a depth camera, our shadow is generated by a model with an angled virtual sun light and detects the nearest point as a pointer. Position of the shadow becomes higher when user walks closer, which conveys the notion of correct distance to control the pointer and offers accessibility to the higher area of the display.

The proliferation of workplace multimedia collaboration applications has meant on one hand more opportunities for group work but on the other more data locked away in proprietary interfaces. We are developing new tools to capture and access multimedia content from any source. In this demo, we focus primarily on new methods that allow users to rapidly reconstitute, enhance, and share document-based information.

Adapting to personal needs and supporting correct posture are important in physiotherapy training. In this demo, we show a dual screen application (handheld and TV) that allows patients to view hypervideo training programs. Designed to guide their daily exercises, these programs can be adapted to daily needs. The dual screen concept offers the positional flexibility missing in single screen solutions.

Dual screen concepts for hypervideo-based physiotherapy training are important in healthcare settings, but existing applications often cannot be adapted to personal needs and do not support correct posture. In this paper, we describe the design and implementation of a dual screen application (handheld and TV) that allows patients to view hypervideos designed to help them correctly perform their exercises. This approach lets patients adapt their training to their daily needs and their overall training progress. We evaluated this prototypical implementation in a user test with post-operative care prostate cancer patients. From our results, we derived design recommendations for dual screen physical training hypervideo applications.

Different systems exist for the creation of hypervideos nowadays. However, the creation of the video scenes which are put together to a hypervideo is a tedious and time consuming job. Then again huge video databases like YouTube exist which already provide rich sources of video materials. Yet it is not allowed to download and re-purpose the videos from there legally, which requires a solution to link whole videos or parts of videos and play them from the platform in an embedded player. This work presents the SIVA Web Producer, a Chrome extension for the creation of hypervideos consisting of scenes from YouTube videos. After creating a project, the Chrome extension allows to import YouTube videos or parts thereof as video clips. These can than be linked to a scene graph. A preview is provided and finalized videos can be published on the SIVA Web Portal.

In this paper we describe DocuGram, a novel tool to capture and share documents from any application. As users scroll through pages of their document inside the native application (Word, Google Docs, web browser), the system captures and analyses in real-time the video frames and reconstitutes the original document pages into an easy to view HTML-based representation. In addition to regenerating the document pages, a DocuGram also includes the interactions users had over them, e.g. mouse motions and voice comments. A DocuGram acts as a modern copy machine, allowing users to copy and share any document from any application.

Most teleconferencing tools treat users in distributed meetings monolithically: all participants are meant to be connected to one another in the more-or-less the same manner. In reality, though, people connect to meetings in all manner of different contexts, sometimes sitting in front of a laptop or tablet giving their full attention, but at other times mobile and involved in other tasks or as a liminal participant in a larger group meeting. In this paper we present the design and evaluation of two applications, Penny and MeetingMate, designed to help users in non-standard contexts participate in meetings.

The abundance of data posted to Twitter enables companies to extract useful information, such as Twitter users who are dissatisfied with a product. We endeavor to determine which Twitter users are potential customers for companies and would be receptive to product recommendations through the language they use in tweets after mentioning a product of interest. With Twitter's API, we collected tweets from users who tweeted about mobile devices or cameras. An expert annotator determined whether each tweet was relevant to customer purchase behavior and whether a user, based on their tweets, eventually bought the product. For the relevance task, among four models, a feed-forward neural network yielded the best cross-validation accuracy of over 80% per product. For customer purchase prediction of a product, we observed improved performance with the use of sequential input of tweets to recurrent models, with an LSTM model being best; we also observed the use of relevance predictions in our model to be more effective with less powerful RNNs and on more difficult tasks.

Two related challenges with current teleoperated robotic systems are lack of peripheral vision and awareness, and difficulty or tedium of navigating through remote spaces. We address these challenges by providing an interface with a focus plus context (F+C) view of the robot location, and where the user can navigate simply by looking where they want to go, and clicking or drawing a path on the view to indicate the desired trajectory or destination. The F+C view provides an undistorted, perspectively correct central region surrounded by a wide field of view peripheral portion, and avoids the need for separate views. The navigation method is direct and intuitive in comparison to keyboard or joystick based navigation, which require the user to be in a control loop as the robot moves. Both the F+C views and the direct click navigation were evaluated in a preliminary user study.

Mobile Telepresence Robots (MTR) are an emerging technology that extend the functionality of telepresence systems by adding mobility. MTRs nowadays, however, rely on stationary imaging systems such as a single narrow-view camera for vision, which can lead to reduced operator performance due to view-related deficiencies in situational awareness. We therefore developed an improved imaging and viewing platform that allows immersive telepresence using a Head Mounted Device (HMD) with head-tracked mono and stereoscopic video. Using a remote collaboration task to ground our research, we examine the effectiveness head-tracked HMD systems in comparison to a baseline monitor-based system. We performed a user study where participants were divided into three groups: fixed camera monitor-based baseline condition (without HMD), HMD with head-tracked 2D camera and HMD with head-tracked stereo camera. Results showed the use of HMD reduces task error rates and improves perceived collaborative success and quality of view, compared to the baseline condition. No major difference was found, however, between stereo and 2D camera conditions for participants wearing an HMD.

Social media offers potential opportunities for businesses to extract business intelligence. This paper presents Tweetviz, an interactive tool to help businesses extract actionable information from a large set of noisy Twitter messages. Tweetviz visualizes tweet sentiment of business locations, identifies other business venues that Twitter users visit, and estimates some simple demographics of the Twitter users frequenting a business. A user study to evaluate the system's ability indicates that Tweetviz can provide an overview of a business's issues and sentiment as well as information aiding users in creating customer profiles.