Papers by Dr.Abhishek Sharma
2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015
This paper proposes a learning-based approach to scene parsing inspired by the deep Recursive Con... more This paper proposes a learning-based approach to scene parsing inspired by the deep Recursive Context Propagation Network (RCPN). RCPN is a deep feed-forward neural network that utilizes the contextual information from the entire image, through bottom-up followed by top-down context propagation via random binary parse trees. This improves the feature representation of every super-pixel in the image for better classification into semantic categories. We analyze RCPN and propose two novel contributions to further improve the model. We first analyze the learning of RCPN parameters and discover the presence of bypass error paths in the computation graph of RCPN that can hinder contextual propagation. We propose to tackle this problem by including the classification loss of the internal nodes of the random parse trees in the original RCPN loss function. Secondly, we use an MRF on the parse tree nodes to model the hierarchical dependency present in the output. Both modifications provide performance boosts over the original RCPN and the new system achieves state-of-the-art performance on Stanford Background, SIFT-Flow and Daimler urban datasets.
2010 Annual IEEE India Conference (INDICON), 2010
Several License Plate Recognition systems have been developed in the past. Our objective is to de... more Several License Plate Recognition systems have been developed in the past. Our objective is to design a system implemented on a standard camera-equipped mobile phone, capable of recognising vehicle license number. As a first step towards it we propose a license plate text segmentation approach that is robust to various lighting conditions, complex background owing to dirty or rusted LP and non-convential fonts. In the Indian scenario, some vehicle owners choose to write their vehicle number plates in regional languages. Since our method does not rely on language-specific features, it is therefore capable of segmenting license number written in different languages. Using color connected component labeling, stroke width and text heuristics we perform the task of accurately segmenting the number from the license plate. Experiments carried out on Indian vehicle license plate (LP) images acquired using a camera-equipped cellphone shows that our system peforms well on different LP images some with different types of degradations. OCR evaluation on the extracted LP number text with the proposed method has an accuracy of 98.86%.
2010 IEEE International Conference on Acoustics, Speech and Signal Processing, 2010
In this paper we present a novel scheme for unstructured audio scene classification that possesse... more In this paper we present a novel scheme for unstructured audio scene classification that possesses three highly desirable and powerful features: autonomy, scalability, and robustness. Our scheme is based on our recently introduced machine learning algorithm called Simultaneous Temporal And Contextual Splitting (STACS) that discovers the appropriate number of states and efficiently learns accurate Hidden Markov Model (HMM) parameters for the given data. STACS-based algorithms train HMMs up to five times faster than Baum-Welch, avoid the overfitting problem commonly encountered in learning large state-space HMMs using Expectation Maximization (EM) methods such as Baum-Welch, and achieve superior classification results on a very diverse dataset with minimal pre-processing. Furthermore, our scheme has proven to be highly effective for building real-world applications and has been integrated into a commercial surveillance system as an event detection component.
2010 IEEE Fourth International Conference on Semantic Computing, 2010
Advances in biomedical technology and research have resulted in a large number of research findin... more Advances in biomedical technology and research have resulted in a large number of research findings, which are primarily published in unstructured text such as journal articles. Text mining techniques have been thus employed to extract knowledge from such data. In this article we focus on the task of identifying and extracting relations between bioentities such as green tea and breast cancer. Unlike previous work that employs heuristics such as co-occurrence patterns and handcrafted syntactic rules, we propose a verb-centric algorithm. This algorithm identifies and extracts the main verb(s) in a sentence; therefore, it does not require the usage of predefined rules or patterns. Using the main verb(s) it then extracts the two involved entities of a relationship. The biomedical entities are identified using a dependence parse tree by applying syntactic and linguistic features such as preposition phrases and semantic role analysis. The proposed verb-centric approach can effectively handle complex sentence structures such as clauses and conjunctive sentences. We evaluate the algorithm on several datasets and achieve an average F-score of 0.905, which is significantly higher than that of previous work.
CVPR 2011, 2011
This paper presents a novel way to perform multi-modal face recognition. We use Partial Least Squ... more This paper presents a novel way to perform multi-modal face recognition. We use Partial Least Squares (PLS) to linearly map images in different modalities to a common linear subspace in which they are highly correlated. PLS has been previously used effectively for feature selection in face recognition. We show both theoretically and experimentally that PLS can be used effectively across modalities. We also formulate a generic intermediate subspace comparison framework for multi-modal recognition. Surprisingly, we achieve high performance using only pixel intensities as features. We experimentally demonstrate the highest published recognition rates on the pose variations in the PIE data set, and also show that PLS can be used to compare sketches to photos, and to compare images taken at different resolutions. 1.1. Related Work There has been a huge amount of prior work on comparing images taken in different modalities, which we
analysis
Nutritional genomics is a new science that studies the relationship between foods (or nutrients),... more Nutritional genomics is a new science that studies the relationship between foods (or nutrients), diseases, and genes. Large amounts of scientific findings have been published in this area, primarily in unstructured text. Moreover, given a pair of entities, different studies can report different findings. It is hence important to obtain a holistic view of the reported relationships. In this article, we describe an information extraction system aiming to reach this goal. The system integrates natural language processing techniques, domain ontology, statistical, and machine learning methods. It consists of four main modules: (1) entity extraction, which recognizes and extracts five types of entities: foods, chemicals (or nutrients), diseases, proteins and genes; (2) relationship extraction, which extracts binary relationships between entities; (3) relationship polarity analysis, which categorizes relationships into three groups: positive, negative, and neutral; and (4) strength analysis, which rates a relationship as weak, medium, or strong. To the best of our knowledge, we are the first to propose to analyze the polarity and strength of a binary relationship. We have evaluated our system using the GENIA corpus and datasets drawn from the MEDLINE database. The first two modules outperform the reported best results with an average Fscore of 0.89 and 0.82, respectively; while the last two also achieve promising results with an accuracy of 0.75-0.84 and ~0.90, respectively.
Computer Vision and Image Understanding, 2012
We propose a novel pose-invariant face recognition approach which we call Discriminant Multiple C... more We propose a novel pose-invariant face recognition approach which we call Discriminant Multiple Coupled Latent Subspace framework. It finds sets of projection directions for different poses such that the projected images of the same subject are maximally correlated in the latent space. Discriminant analysis with artificially simulated pose errors in the latent space makes it robust to small pose errors caused due to a subject's incorrect pose estimation. We do a comparative analysis of three popular learning approaches: Partial Least Squares (PLS), Bilinear Model (BLM) and Canonical Correlational Analysis (CCA) in the proposed coupled latent subspace framework. We also show that using more than two poses simultaneously with CCA results in better performance. We report state-of-the-art results for pose-invariant face recognition on CMU PIE and FERET and comparable results on MultiPIE when using only 4 fiducial points and intensity features.
Systems that could learn by reading would radically change the economics of building large knowle... more Systems that could learn by reading would radically change the economics of building large knowledge bases. This paper describes Learning Reader, a prototype system that extends its knowledge base by reading. Learning Reader consists of three components. The Reader, which converts text into formally represented cases, uses a Direct Memory Access Parser operating over a large knowledge base, derived from ResearchCyc. The Q/A system, which provides a means of quizzing the system on what it has learned, uses focused sets of axioms automatically extracted from the knowledge base for tractability. The Ruminator, which attempts to improve the system's understanding of what it has read by off-line processing, generates questions for itself by several means, including analogies with prior material and automatically constructed generalizations from examples in the KB and its prior reading. We discuss the architecture of the system, how each component works, and some experimental results.
Lecture Notes in Computer Science, 2015
In this paper, we present a novel approach to study and reveal network and protocol information f... more In this paper, we present a novel approach to study and reveal network and protocol information from energy instrumentation in wireless sensor network. Unlike prior approaches which focused on analyzing the aggregate statistics of energy efficiency of a network or a protocol, our approach aims at revealing network protocols, application workloads, and topology information by fine-grained energy instrumentation on the nodes. We design a set of features based on various aspects of energy data and use those features to classify and reveal network activity. Results from experiments on three testbeds indicate that our approach can achieve up to 97% accuracy to identify the routing protocols, and infer the network topology with 98% accuracy.
Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems, 2014
In this poster, we present a novel approach to study and reveal network protocol information from... more In this poster, we present a novel approach to study and reveal network protocol information from radio activities instrumentation in wireless sensor network. Recent studies have analyzed radio activities; however, most of these studies focus on estimating energy consumption, since radio chip usually dominates the energy consumption of nodes. In our work, we analyze radio activities with a different purpose, which aims to reveal network protocols and application workloads by an analysis of fine-grained low level radio activities on the nodes. We design a feature called Radio Awake Length Counter and use it to classify and reveal network activity. Results from experiments on a real world testbed indicate that our approach can achieve up to 97% accuracy to identify the routing protocols, average 85% accuracy to distinguish application workloads.
Cloud operators increasingly need many fine-grained rules to better control individual network fl... more Cloud operators increasingly need many fine-grained rules to better control individual network flows for various management tasks. While previous approaches have advocated placing rules either on hypervisors or switches, we argue that future data centers would benefit from leveraging rule processing capabilities at both for better scalability and performance. In this paper, we propose vCRIB, a virtualized Cloud Rule Information Base that allows operators to freely define different management policies without the need to consider underlying resource constraints. The challenge in our approach is the design of a vCRIB manager that automatically partitions and places rules at both hypervisors and switches to achieve a good trade-off between resource usage and performance.
International Journal of Thermodynamics, 2014
Performance of a gas turbine is mainly depends on the inlet air temperature. The power output of ... more Performance of a gas turbine is mainly depends on the inlet air temperature. The power output of a gas turbine depends on the flow of mass through it. Inlet air cooling increases the power output by taking advantage of the gas turbine's feature of higher mass flow rate when the compressor inlet temperature decreases. This is precisely the reason why on hot days, when air is less dense, power output falls off. A rise of 1°C temperature of inlet air decreases the power output by 1%. In this paper the performance enhancement of gas turbine power plants by cooling the compressor intake air with an evaporative cooler is studied. This study investigated the effect of inlet air cooling system on the performance of an existing gas turbine power plant in Nigeria. The results show that for each 5 o C decrease of inlet air temperature, net output power increases around 5-10% and the first and second law efficiencies increase around 2-5%. It is shown that the amount of this increase is higher when the pressure ratio is high and turbine inlet temperature is low. The results of this study shows that retrofitting of the existing gas turbine plant with inlet air cooling system gives a better system performance and may prove to be an attractive investment opportunity to the Nigeria government and stakeholders of the plant.
Irrigation water pricing: the gap between theory and practice, 2007
This report first assesses the scale of the energy-irrigation nexus in South Asia. This is follow... more This report first assesses the scale of the energy-irrigation nexus in South Asia. This is followed by a section describing what it would take to make a metered tariff regime work, the main comparison being with North China where such a regime does seem to work. The potential for indirect management of the groundwater economy through the specific mechanism of electricity pricing and supply policies is discussed.
The agricultural groundwater revolution: opportunities and threats to development
In serving this mission, IWMI concentrates on the integration of policies, technologies and manag... more In serving this mission, IWMI concentrates on the integration of policies, technologies and management systems to achieve workable solutions to real problems-practical, relevant results in the field of irrigation and water and land resources. The publications in this series cover a wide range of subjects-from computer modeling to experience with water user associations-and vary in content from directly applicable research to more basic studies, on which applied work ultimately depends. Some research reports are narrowly focused, analytical and detailed empirical studies; others are wide-ranging and synthetic overviews of generic problems. Although most of the reports are published by IWMI staff and their collaborators, we welcome contributions from others. Each report is reviewed internally by IWMI's own staff and Fellows, and by external reviewers. The reports are published and distributed both in hard copy and electronically (www.iwmi.org) and where possible all data and analyses will be available as separate downloadable files. Reports may be copied freely and cited with due acknowledgment.
Lecture Notes in Computer Science, 2006
Observing systems facilitate scientific studies by instrumenting the real world and collecting co... more Observing systems facilitate scientific studies by instrumenting the real world and collecting corresponding measurements, with the aim of detecting and tracking phenomena of interest. A wide range of critical environmental monitoring objectives in resource management, environmental protection, and public health all require distributed observing systems. The goal of such systems is to help scientists verify or falsify hypotheses with useful samples taken by the stationary and mobile units, as well as to analyze data autonomously to discover interesting trends or alarming conditions. In our project, we focus on a class of observing systems which are embedded into the environment, consist of stationary and mobile sensors, and react to collected observations by reconfiguring the system and adapting which observations are collected next. In this paper, we give an overview of our project in the context of a marine biology application.
Lecture Notes in Computer Science, 2007
Observing systems facilitate scientific studies by instrumenting the real world and collecting co... more Observing systems facilitate scientific studies by instrumenting the real world and collecting corresponding measurements, with the aim of detecting and tracking phenomena of interest. Our AMBROSia project focuses on a class of observing systems which are embedded into the environment, consist of stationary and mobile sensors, and react to collected observations by reconfiguring the system and adapting which observations are collected next. In this paper, we report on recent research directions and corresponding results in the context of AMBROSia.
Performance Evaluation, 2010
Wireless sensor systems aid scientific studies by instrumenting the real world and collecting mea... more Wireless sensor systems aid scientific studies by instrumenting the real world and collecting measurements. Given the large volume of measurements collected by sensor systems, one problem arises-an automated approach to identifying the "interesting" parts of these data sets, or anomaly detection. A good anomaly detection methodology should be able to accurately identify many types of anomalies, be robust, require relatively little resources, and perform detection in (near) real-time. Thus, in this paper we focus on an approach to online anomaly detection in measurements collected by sensor systems, where our evaluation, using real-world datasets, shows that our approach is accurate (it detects over 90% of the anomalies with few false positives), works well over a range of parameter choices, and has a small (CPU, memory) footprint.
arXiv (Cornell University), Feb 9, 2019
Scientists deploy environmental monitoring networks to discover previously unobservable phenomena... more Scientists deploy environmental monitoring networks to discover previously unobservable phenomena and quantify subtle spatial and temporal differences in the physical quantities they measure. Our experience, shared by others, has shown that measurements gathered by such networks are perturbed by sensor faults. In response, multiple fault detection techniques have been proposed in the literature. However, in this paper we argue that these techniques may mis-classify events (e.g. rain events for soil moisture measurements) as faults, potentially discarding the most interesting measurements. We support this argument by applying two commonly used fault detection techniques on data collected from a soil monitoring network. Our results show that in this case, up to 45% of the event measurements are misclassified as faults. Furthermore, tuning the fault detection algorithms to avoid event misclassification, causes them to miss the majority of actual faults. In addition to exposing the tension between fault and event detection, our findings motivate the need to develop novel fault detection mechanisms which incorporate knowledge of the underlying events and are customized to the sensing modality they monitor.
The Laugh Machine project aims at endowing virtual agents with the capability to laugh naturally,... more The Laugh Machine project aims at endowing virtual agents with the capability to laugh naturally, at the right moment and with the correct intensity, when interacting with human participants. In this report we present the technical development and evaluation of such an agent in one specific scenario: watching TV along with a participant. The agent must be able to react to both, the video and the participant's behaviour. A full processing chain has been implemented, integrating components to sense the human behaviours, decide when and how to laugh and, finally, synthesize audiovisual laughter animations. The system was evaluated in its capability to enhance the affective experience of naive participants, with the help of pre and post-experiment questionnaires. Three interaction conditions have been compared: laughter-enabled or not, reacting to the participant's behaviour or not. Preliminary results (the number of experiments is currently to small to obtain statistically significant differences) show that the interactive, laughter-enabled agent is positively perceived and is increasing the emotional dimension of the experiment.
Computing-as-a-service has been evolving steadily. Today, private clouds (e.g., Google's inte... more Computing-as-a-service has been evolving steadily. Today, private clouds (e.g., Google's internal shared computing cluster) as well as public clouds (e.g., Amazon's web services (AWS), Microsoft's Azure) provide computing abstractions at various levels: bare virtual machines, specialized languages and runtimes (e.g., for massively-parallel data processing---MapReduce, Dryad), web services. For example, Amazon offers bare virtual machines as well as MapReduce clusters.
Uploads
Papers by Dr.Abhishek Sharma