Current state-of-the-art object detectors can have significant performance drop when deployed in the wild due to domain gaps with training data. Unsupervised Domain Adaptation (UDA) is a promising approach to adapt detectors for new domains/environments without any expensive label cost. Previous mainstream UDA works for object detection usually focused on image-level and/or feature-level adaptation by using adversarial learning methods. In this work, we show that such adversarial-based methods can only reduce domain style gap, but cannot address the domain content gap that is also important for object detectors. To overcome this limitation, we propose the SC-UDA framework to concurrently reduce both gaps: We propose fine-grained domain style transfer to reduce the style gaps with finer image details preserved for detecting small objects; Then we leverage the pseudo label-based self-training to reduce content gaps; To address pseudo label error accumulation during self-training, novel optimizations are proposed, including uncertainty-based pseudo labeling and imbalanced mini-batch sampling strategy. Experiment results show that our approach consistently outperforms prior state-of-the-art methods (up to 8.6%, 2.7% and 2.5% mAP on three UDA benchmarks).

This paper gives an overview of our ongoing work on the design space exploration of efficient deep neural networks (DNNs), specifically on the novel optimization perspectives that past work have mainly overlooked. We cover two complementary aspects of efficient DNN design: (1) static architecture design efficiency and (2) dynamic model execution efficiency. In the static architecture design, one of the major challenges of NAS is the low search efficiency. Different with current mainstream efficient search algorithm optimization, we identify the new perspective in efficient search space design. In the dynamic model execution, current major optimization methods still target at the model structure redundancy, e.g., weight/filter pruning, connection pruning, etc. We instead identify the new dimension of DNN feature map redundancy. By showcasing such new perspectives, further advantages could be potentially attained by integrating both current optimizations and our new perspectives.

In order to match shoppers with desired products and provide personalized promotions, whether in online or offline shopping worlds, it is critical to model both consumer preferences and price sensitivities simultaneously. Personalized preferences have been thoroughly studied in the field of recommender systems, though price (and price sensitivity) has received relatively little attention. At the same time, price sensitivity has been richly explored in the area of economics, though typically not in the context of developing scalable, working systems to generate recommendations. In this study, we seek to bridge the gap between large-scale recommender systems and established consumer theories from economics, and propose a nested feature-based matrix factorization framework to model both preferences and price sensitivities. Quantitative and qualitative results indicate the proposed personalized, interpretable and scalable framework is capable of providing satisfying recommendations (on two datasets of grocery transactions) and can be applied to obtain economic insights into consumer behavior.

We present the results, experiences and lessons learned from comparing a diverse set of technical approaches to indoor localization during the 2014 Microsoft Indoor Localization Competition. 22 different solutions to indoor localization from different teams around the world were put to test in the same unfamiliar space over the course of 2 days, allowing us to directly compare the accuracy and overhead of various technologies. In this paper, we provide a detailed analysis of the evaluation study's results, discuss the current state-of-the-art in indoor localization, and highlight the areas that, based on our experience from organizing this event, need to be improved to enable the adoption of indoor location services.

Trending search topics cause unpredictable query load spikes that hurt the end-user search experience, particularly the mobile one, by introducing longer delays. To understand how trending search topics are formed and evolve over time, we analyze 21 million queries submitted during periods where popular events caused search query volume spikes. Based on our findings, we design and evaluate PocketTrend, a system that automatically detects trending topics in real time, identifies the search content associated to the topics, and then intelligently pushes this content to users in a timely manner. In that way, PocketTrend enables a client-side search engine that can instantly answer user queries related to trending events, while at the same time reducing the impact of these trends on the datacenter workload. Our results, using real mobile search logs, show that in the presence of a trending event, up to 13-17% of the overall search traffic can be eliminated from the datacenter, with as many as 19% of all users benefiting from PocketTrend.

Local search users today decide what business to visit solely based on distance information, and business ratings that can be sparse or stale. We believe that when users search for local businesses, such as bars or restaurants, they need to know more about the ambience of each business, such as how crowded it is, how loud and of what type the music it plays is, as well as how loud the human chatter in the business is. Unfortunately, this information doesn't exist today. In this paper, we propose to automatically crowdsource such rich, local business ambience metadata through real user check-in events. Every time a user checks into a business, the phone is in user's hands, and the phone's sensors can sense the business environment. We leverage the phone's microphone during this time to infer the occupancy and human chatter levels, the music type, as well as the music and noise levels in the business. As people check-in to businesses throughout the day, business metadata can be automatically updated over time, enabling a new generation of local search experience. Using approximately 150 audio traces collected from real businesses of various types over a period of 3 months, we show that by properly extracting the temporal and frequency signatures of the audio signal, it is feasible to train models that can simultaneously infer occupancy, human chatter, music, and noise levels in a business, with higher than 79% accuracy.

The major challenge for accurate fingerprint-based indoor localization is the design of robust and discriminative wireless signatures. Even though WiFi received signal strength indicator (RSSI) signatures are widely available indoors, they vary significantly over time and are susceptible to human presence, multipath, and fading due to the high operating frequency. To overcome these limitations, we propose to use FM broadcast radio signals for robust indoor fingerprinting. Because of the lower frequency, FM signals are less susceptible to human presence, multipath, and fading, they exhibit exceptional indoor penetration, and according to our experimental study they vary less over time when compared to WiFi signals. In this paper, we demonstrate through a detailed experimental study in three different buildings across the US, that FM radio signal RSSI values can be used to achieve room-level indoor localization with similar or better accuracy to the one achieved by WiFi signals. Furthermore, we propose to use additional signal quality indicators at the physical layer (i.e., SNR, multipath, etc.) to augment the wireless signature, and show that localization accuracy can be further improved by more than 5 percent. More importantly, we experimentally demonstrate that the localization errors of FM and WiFi signals are independent. When FM and WiFi signals are combined to generate wireless fingerprints, the localization accuracy increases as much as 83 percent (when accounting for wireless signal temporal variations) compared to when WiFi RSSI only is used as a signature.

The high network latencies and limited battery life of mobile phones can make mobile web browsing a frustrating experience. In prior work, we proposed trading memory capacity for lower web access latency and a more convenient data transfer schedule from an energy perspective by prefetching slowly-changing data (search queries and results) nightly, when the phone is charging. However, most web content is intrinsically much more dynamic and may be updated multiple times a day, thus eliminating the effectiveness of periodic updates.

The major challenge for accurate fingerprint-based indoor localization is the design of robust and discriminative wireless signatures. Even though WiFi RSSI signatures are widely available indoors, they vary significantly over time and are susceptible to human presence, multipath, and fading due to the high operating frequency. To overcome these limitations, we propose to use FM broadcast radio signals for robust indoor fingerprinting. Because of the lower frequency, FM signals are less susceptible to human presence, multipath and fading, they exhibit exceptional indoor penetration, and according to our experimental study they vary less over time when compared to WiFi signals. In this work, we demonstrate through a detailed experimental study in 3 different buildings across the US, that FM radio signal RSSI values can be used to achieve room-level indoor localization with similar or better accuracy to the one achieved by WiFi signals. Furthermore, we propose to use additional signal quality indicators at the physical layer (i.e., SNR, multipath etc.) to augment the wireless signature, and show that localization accuracy can be further improved by more than 5%. More importantly, we experimentally demonstrate that the localization errors of FM andWiFi signals are independent. When FM and WiFi signals are combined to generate wireless fingerprints, the localization accuracy increases as much as 83% (when accounting for wireless signal temporal variations) compared to when WiFi RSSI only is used as a signature.

Users increasingly rely on their mobile devices to search local entities, typically businesses, while on the go. Even though recent work has recognized that the ranking signals in mobile local search (e.g., distance and customer rating score of a business) are quite different from general Web search, they have mostly treated these signals as a black-box to extract very basic features (e.g., raw distance values and rating scores) without going inside the signals to understand how exactly they affect the relevance of a business. However, as it has been demonstrated in the development of general information retrieval models, it is critical to explore the underlying behaviors/heuristics of a ranking signal to design more effective ranking features.

Cloud services accessed through mobile devices suffer from high network access latencies and are constrained by energy budgets dictated by the devices' batteries. Radio and battery technologies will improve over time, but are still expected to be the bottlenecks in future systems. Non-volatile memories (NVM), however, may continue experiencing significant and steady improvements in density for at least ten more years. In this paper, we propose to leverage the abundance in memory capacity of mobile devices to mitigate latency and energy issues when accessing cloud services.

Users increasingly rely on their mobile devices to search, locate and discover places and activities around them while on the go. Their decision process is driven by the information displayed on their devices and their current context (e.g. traffic, driving or walking etc.). Even though recent research efforts have already examined and demonstrated how different context parameters such as weather, time and personal preferences affect the way mobile users click on local businesses, little has been done to study how the location of the user affects the click behavior. In this paper we follow a data-driven methodology where we analyze approximately 2 million local search queries submitted by users across the US, to visualize and quantify how differently mobile users click across locations. Based on the data analysis, we propose new location-aware features for improving local search click prediction and quantify their performance on real user query traces. Motivated by the results, we implement and evaluate a data-driven technique where local search models at different levels of location granularity (e.g. city, state, and country levels) are combined together at run-time to further improve click prediction accuracy. By applying the location-aware features and the multiple models at different levels of location granularity on real user query streams from a major, commercially available search engine, we achieve anywhere from 5% to 47% higher Precision than a single click prediction model across the US can achieve.

The signals used for ranking in local search are very different from web search: in addition to (textual) relevance, measures of (geographic) distance between the user and the search result, as well as measures of popularity of the result are important for effective ranking. Depending on the query and search result, different ways to quantify these factors exist -- for example, it is possible to use customer ratings to quantify the popularity of restaurants, whereas different measures are more appropriate for other types of businesses. Hence, our approach is to capture the different notions of distance/popularity relevant via a number of external data sources (e.g., logs of customer ratings, driving-direction requests, or site accesses).

Local search engines are very popular but limited. We present Hapori, a next-generation local search technology for mobile phones that not only takes into account location in the search query but richer context such as the time, weather and the activity of the user. Hapori also builds behavioral models of users and exploits the similarity between users to tailor search results to personal tastes rather than provide static geo-driven points of interest. We discuss the design, implementation and evaluation of the Hapori framework which combines data mining, information preserving embedding and distance metric learning to address the challenge of creating efficient multidimensional models from context-rich local search logs. Our experimental results using 80,000 queries extracted from search logs show that contextual and behavioral similarity information can improve the relevance of local search results by up to ten times when compared to the results currently provided by commercially available search engine technology.

Using data from 43 users across two platforms, we present a detailed look at smartphone traffic. We find that browsing contributes over half of the traffic, while each of email, media, and maps contribute roughly 10%. We also find that the overhead of lower layer protocols is high because of small transfer sizes. For half of the transfers that use transport-level security, header bytes correspond to 40% of the total. We show that while packet loss is the main factor that limits the throughput of smartphone traffic, larger send buffers at Internet servers can improve the throughput of a quarter of the transfers. Finally, by studying the interaction between smartphone traffic and the radio power management policy, we find that the power consumption of the radio can be reduced by 35% with minimal impact on the performance of packet exchanges.

Although mobile phones are ideal platforms for continuous human centric sensing, the state of the art phone architectures today have not been designed to support continuous sensing applications. Currently, sampling and processing sensor data on the phone requires the main processor and associated components to be continuously on, creating a large energy overhead that can severely impact the battery lifetime of the phone. We will demonstrate Little Rock, a novel sensing architecture for mobile phones, where sampling and, when possible, processing of sensor data is offloaded to a dedicated low-power processor. This approach enables the phone to perform continuous sensing three orders of magnitude more energy efficiently compared to the normal approaches.

Using detailed traces from 255 users, we conduct a comprehensive study of smartphone use. We characterize intentional user activities -- interactions with the device and the applications used -- and the impact of those activities on network and energy usage. We find immense diversity among users. Along all aspects that we study, users differ by one or more orders of magnitude. For instance, the average number of interactions per day varies from 10 to 200, and the average amount of data received per day varies from 1 to 1000 MB. This level of diversity suggests that mechanisms to improve user experience or energy consumption will be more effective if they learn and adapt to user behavior. We find that qualitative similarities exist among users that facilitate the task of learning user behavior. For instance, the relative application popularity for can be modeled using an exponential distribution, with different distribution parameters for different users. We demonstrate the value of adapting to user behavior in the context of a mechanism to predict future energy drain. The 90th percentile error with adaptation is less than half compared to predictions based on average behavior across users.

Mobile applications often need location data, to update locally relevant information and adapt the device context. While most smartphones do include a GPS receiver, it's frequent use is restricted due to high battery drain. We design and prototype an adaptive location service for mobile devices, a-Loc, that helps reduce this battery drain. Our design is based on the observation that the required location accuracy varies with location, and hence lower energy and lower accuracy localization methods, such as those based on WiFi and cell-tower triangulation, can sometimes be used. Our method automatically determines the dynamic accuracy requirement for mobile search-based applications. As the user moves, both the accuracy requirements and the location sensor errors change. A-Loc continually tunes the energy expenditure to meet the changing accuracy requirements using the available sensors. A Bayesian estimation framework is used to model user location and sensor errors. Experiments are performed with Android G1 and AT&T Tilt phones, on paths that include outdoor and indoor locations, using war-driving data from Google and Microsoft. The experiments show that a-Loc not only provides significant energy savings, but also improves the accuracy achieved, because it uses multiple sensors.

The extraction of temporal characteristics from sensor data streams can reveal important properties about the sensed events. Knowledge of temporal characteristics in applications where sensed events tend to periodically repeat, can provide a great deal of information towards identifying patterns, building models and using the timing information to actuate and provide services. In this paper we outline a methodology for extracting the temporal properties, in terms of start time and duration, of sensor data streams that can be used in applications such as human, habitat, environmental and traffic monitoring where sensed events repeat over a time window. Its application is demonstrated on a 30-day dataset collected from one of our assisted living sensor network deployments.

Osteoarthritis (OA) is a major cause of morbidity worldwide, representing the most common form of arthritis. The radiographic assessment of OA-severity is mainly relied on qualitative criteria, evaluating structural alterations of the joint. In the present study a computer-based image analysis method was developed for the grading of hip OA-severity from radiographic images. The sample of the study comprised 64 hips (18 normal, 46 osteoarthritic), corresponding to 32 unilateral and bilateral hip-OA patients. Two experienced orthopaedists assessed OA-severity from pelvic radiographs, employing the Kellgren and Lawrence (KL) grading scale. Accordingly, 3 KL-based OA-severity categories were formed: (i) “Normal / Doubtful”, (ii) “Mild / Moderate”, and “Severe”. After radiographs digitization their contrast was enhanced by means of the Contrast Limited Adaptive Histogram Equalization method. Employing custom developed algorithms: (i) 64 ROIs, corresponding to patients’ Hip Joint Spaces (HJSs), were determined on the processed radiographs, and (ii) the Radial Distance Signature (RDS) of each HJS-ROI was generated, as the sequence of the Euclidean radial distances between the “centre of mass (centroid)” and each point of the HJS-ROI contour. The generated RDS was subject to the Discrete Wavelet Transform (Coiflet1 wavelet, Level 2 decomposition). Statistical measures of the generated wavelet coefficients were used for the formation of feature vectors, representative of the HJS-ROIs. These vectors were involved in the design of a grading system, based on the Bayes classifier, which was used for the discrimination between: (i) normal and OA hips, and (ii) hips of “Mild / Moderate” and “Severe” OA. The classification accuracy achieved regarding the discrimination between normal and OA hips was 95.3%, while the relevant score for the characterization of hips as of “Mild / Moderate” or “Severe” OA was 91.3%. The proposed system could be of value for the management of hip OA patient.

Researchers have portrayed an increasing effort towards providing formal computational frameworks to consolidate the plethora of notions and relations used in the medical domain. Despite the fact that there are many reasons for this, the need for standardization of protocols and terminology is critical, not only for the provision of uniform levels of health care, but also to facilitate medical science research. In the domain of skin lesions, the variability of semantic features contained within images is major among the barriers to the medical understanding of the symptoms and development of early skin cancers. Such variability is acknowledged across specialist fields of medicine, motivating standardization of terminologies for reporting medical practice. The desideratum of making these standards machine-readable has led to their formalization in the form of ontologies. Ontologies are computational artifacts designed to provide semantic representations of a particular domain of interest. Such a representation can be encoded and reused, allowing navigation of the key concepts recorded and retrieval of information indexed against it. This fact bridges the required standardization gap by offering a set of labeling options to record observations and events encountered by medical experts. Given the twin goals of ontologies - representation and standardization - the present paper deals with the creation of an ontology for skin lesion images, encoded using Description Logic in OWL format, which can be used for decision support systems that operate on a classification basis. The declarative framework within which ontologies are encoded can transcend any specific application context and help dermatologists perform semantic inference on skin lesion images. Furthermore, extensions are easier to be accomplished and new classes can be straightforwardly incorporated.

Embedded systems with heterogeneous processors extend the energy/timing trade-off flexibility and provide the opportunity to fine tune resource utilization for particular applications. In this paper, we present a resource model that considers the time and energy costs of run-time mode switching, which considerably improves the accuracy of existing models. Given an application, the software partitioning problem then becomes an optimization over energy cost given deadline constraints, which can be formulate as an integer linear programming (ILP) problem. We apply the resource modeling and software partitioning techniques to a multimodule embedded sensing device, the mPlatform, and present a case study of configuring the platform for a real-time sound source localization application on a stack of MSP430 and ARM7 processor based sensing and processing boards.

In this demonstration we present the BehaviorScope, a system for interpreting human activity patterns using a sensor network and its application to elder monitoring in assisted living. The BehaviorScope provides a runtime, user-programmable framework that processes streams of timestamped sensor data along with prior context information to infer activities and generate appropriate notifications. Human activities are described in high-level scripts that are directly mapped to a hierarchy of probabilistic grammars that parse low-level sensor measurements into high-level distinguishable activities. Activities of interest are pre-programmed into a specification that is used by the system to interpret the incoming sensor data stream. The system interprets the activities to generate summaries and other triggered notifications that are propagated to stakeholders via email and cell-phone text messages.

We study the problem of concurrently supporting multiple radios with different capabilities and interfaces on a single sensor node platform. Through a detailed experimental study on hardware multi-radio platforms, using the two representative radio technologies 802.15.4 and 802.11, we identify bottlenecks and design tradeoffs that are usually overlooked and that, as we show, have a significant impact on the sensor network's performance and energy efficiency. Our findings are threefold. We show that a proper pairing of processor and radio is crucial for taking the full advantage of the energy efficiency of higher bandwidth radios. The processor/radio pairing affects the energy balance of a sensor node, thus making the design of dynamic switching among multiple radios more challenging. Second, we demonstrate and quantify the impact of network traffic on energy consumption of a sensor node while varying network parameters, and illustrate the deficiency of existing energy-optimizing protocols. Our results indicate that by properly adjusting network parameters, such as packet size and transmission period, energy savings of up to 50% can be achieved under heavy network traffic conditions when a CSMA-based MAC is used. We conclude by presenting a set of guidelines for designing and implementing energy efficient multi-radio platforms.

This paper presents BScope, a new system for interpreting human activity patterns using a sensor network. BScope provides a runtime, user-programmable framework that processes streams of timestamped sensor data along with prior context information to infer activities and generate appropriate notifications. The users of the system are able to describe human activities with high-level scripts that are directly mapped to hierarchical probabilistic grammars used to parse low-level sensor measurements into high-level distinguishable activities. Our approach is presented, though not limited, in the context of an assisted living application in which a small, privacy-preserving camera sensor network of five nodes is used to monitor activity in the entire house over a period of 25 days. Privacy is preserved by the fact that camera sensors only provide discrete high-level features, such as motion information in the form of image locations, and not actual images. In this deployment, our primary sensing modality is a distributed array of image sensors with wide-angle lenses that observe people's locations in the house during the course of the day. We demonstrate that our system can successfully generate summaries of everyday activities and trigger notifications at runtime by using more than 1.3 million location measurements acquired through our real home deployment.

We present mPlatform, a new reconfigurable modular sensornet platform that enables real-time processing on multiple heterogeneous processors. At the heart of the mPlatform is a scalable high-performance communication bus connecting the different modules of a node, allowing time-critical data to be shared without delay and supporting reconfigurability at the hardware level. Furthermore, the bus allows components of an application to span across different processors/modules without incurring much overhead, thus easing the program development and supporting software reconfigurability. We describe the communication architecture, protocol, and hardware configuration, and the implementation in a low power, high speed complex programmable logic device (CPLD). An asynchronous interface decouples the local processor of each module from the bus, allowing the bus to operate at the maximum desired speed while letting the processors focus on their real time tasks such as data collection and processing. Extensive experiments on the mPlatform prototype have validated the scalability of the communication architecture, and the high speed, reconfigurable inter-module communication that is achieved at the expense of a small increase in the power consumption. Finally, we demonstrate a real-time sound source localization application on the mPlatform, with four channels of acoustic data acquisition, FFT, and sound classification, that otherwise would be infeasible using traditional buses such as I2C.

Camera sensors constitute an information rich sensing modality with many potential applications in sensor networks. Their effectiveness in a sensor network setting however greatly relies on their ability to calibrate with respect to each other, and other sensors in the field. This paper examines node localization and camera calibration using the shared field of view of camera pairs. Using a new distributed camera sensor network we compare two approaches from computer vision and propose an algorithm that combines a sparse set of distance measurements with image information to accurately localize nodes in 3D. Our algorithms are evaluated using a network of iMote2 nodes equipped with COTS camera modules. The sensor nodes identify themselves to cameras using modulated LED emissions. Our indoor experiments yielded a 2-7cm error in a 6x6m room. Our outdoor experiments in a 30x30m field resulted in errors 20-80cm, depending on the method used.

The ability of a sensor network to parse out observable activities into a set of distinguishable actions is a powerful feature that can potentially enable many applications of sensor networks to everyday life situations. In this paper we introduce a framework that uses a hierarchy of Probabilistic Context Free Grammars (PCFGs) to perform such parsing. The power of the framework comes from the hierarchical organization of grammars that allows the use of simple local sensor measurements for reasoning about more macroscopic behaviors. Our presentation describes how to use a set of phonemes to construct grammars and how to achieve distributed operation using a messaging model. The proposed framework is flexible. It can be mapped to a network hierarchy or can be applied sequentially and across the network to infer behaviors as they unfold in space and time. We demonstrate this functionality by inferring simple motion patterns using a sequence of simple direction vectors obtained from our camera sensor network testbed.

Although imaging is an information-rich sensing modality, the use of cameras in sensor networks is very often prohibited by factors such as power, computation cost, storage, communication bandwidth and privacy. In this paper we consider information selective and privacy-preserving address-event imagers for sensor networks. Instead of providing full images with a high degree of redundancy, our efforts in the design of these imagers specialize on selecting a handful of features from a scene and outputting these features in address-event representation. In this paper we present our initial results in modeling and evaluating address-event sensors in the context of sensor networks. Using three different platforms that we have developed, we illustrate how to model address-event cameras and how to build an emulator using these models. We also present a lightweight classification scheme to illustrate the computational advantages of address-event sensors. The paper concludes with an evaluation of the classification algorithm and a feasibility study of using COTS components to emulate address-event inside a sensor network.

This paper describes the XYZ, a new open-source sensing platform specifically designed to support our experimental research in mobile sensor networks. The XYZ node is designed around the OKI ML67Q500x ARM THUMB Microprocessor and the IEEE 802.15.4 compliant CC2420 radio from Chipcon. Its new features include support for two different CPU sleep modes and a long-term ultra low power sleep mode for the entire node. This allows the XYZ and its peripheral boards to transition into deep sleep for extended time intervals. To support mobility hardware control and computations, XYZ supports a wide dynamic range and power options. In low power configuration the node resembles existing small low power nodes. When needed, the node can scale up its resources to perform more powerful computations. Mobility is enabled with an additional accessory board that allows the node to move along a horizontal string. In this paper we provide an overview of the XYZ architecture and provide an insightful power characterization of the different operational modes to allow the users to optimize their platforms for power.

The use of low-cost imagers in sensor network applications is imminent. Small low-power imagers have many potential applications in visual confirmation of events, target recognition, security and entertainment. To make the autonomous operation of imager sensors in such applications possible, smart imager sensors should be able to calibrate their views with respect to other imagers in the network and should become aware of the locations of other non-imager sensors. Such calibration will allow cameras with multiple vantage points and sensors distributed in physical space to jointly undertake collaborative tasks.