188宝金博页面版

A Formal Algebra Implementationfor Distributed Image and Video Stream ProcessingMohamed A. Helala, Ken Q. Pu, Faisal Z. QureshiFaculty of Science, University of Ontario Institute of Technology, Oshawa, ON, Canada{Mohamed.Helala, Ken.Pu, Faisal.Qureshi}@uoit.caABSTRACTWe are interested in building scalable computer vision systems fordistributed processing of big visual data. We apply data stream-ing concepts, namely stream algebra operators, which have beenproven effective in the database literature. The operators collec-tively form an algebra over data streams. The algebra has welldef i ned semantics. It naturally describes online computer visionalgorithms and their feedback control and tuning algorithms.In this work, we present the f i rst implementation of such alge-bra at large scale. Our implementation provides a high level pro-gramming interface for constructing and executing vision workf l owgraphs while hiding the data transfer and concurrency details. Italso allows feedback control and dynamic reconf i guration of visionalgorithms. A case study is discussed showing a streaming work-f l ow for online lane and road boundary detection and describingthe f l exibility and effectiveness of the algebra for building complexdistributed applications.Keywordsstream algebra; workf l ow graphs; programming frameworks; dis-tributed vision processing1. INTRODUCTIONOur ability to create, store, process, share and consume imageand video data has increased manifolds over the last decade. If thecurrent trend is any guide, we can expect this increase to continueunabated. There is also a new found interest in computing plat-forms aimed specif i cally at processing large volumes of image andvideo data—Nvidia (TM) , for example, has come out with a num-ber of GPUs that specialize in visual processing. What is, how-ever, still missing are easy-to-use mechanisms and frameworks thatwill allow us to eff i ciently deploy large-scale visual processing onthe available computational hardware. We f i nd instances of suchframeworks in the text processing domain. Internet scale text pro-cessing, which forms the backbone of the many of the services thatwe use daily, relies on the existence of such frameworks—HadoopMap-Reduce [3], Storm [2], etc.—that are able to maximize the useof available computational resources. The success of these frame-works is due to their ability to “intelligently” respond to changingcomputational environment and data landscape.Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor prof i t or commercial advantage and that copies bear this notice and the full cita-tion on the f i rst page. Copyrights for components of this work owned by others thanACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-publish, to post on servers or to redistribute to lists, requires prior specif i c permissionand/or a fee. Request permissions from permissions@acm.org.ICDSC ’16, September 12-15, 2016, Paris, Francec ? 2016 ACM. ISBN 978-1-4503-4786-0/16/09...$15.00DOI: http://dx.doi.org/10.1145/2967413.2967432Within this backdrop, we are specif i cally interested in streamingframeworks for visual data. Streaming frameworks are designedto deal with continuous arrival of data (in our case, images andvideos). Our interest in streaming frameworks stems for the ob-servation that such frameworks can be used to implement onlinevision algorithms. A number of common computer vision algo-rithms, such as optical f l ow estimation, stereo estimation, mosaicstitching, action recognition, etc., can be seen as online algorithmsoperating upon continuous streams of incoming images and videos.Several online algorithms that deal with image and video streams(henceforth, we will refer to these streams as visual streams) al-ready exist in the literature: streaming hierarchical video segmenta-tion [21], human body segmentation from video stream [12], photostream alignment [18], and inference of storylines from web photostreams [22, 16].Streamingalgorithmsareoftenrepresentedaspipelines(orwork-f l ows). The analogy is a sound one; raw data enters at one end andprocessed information appears at the other end of a pipeline. Theconcept of pipelines f i rst appeared in the database literature, wherethe notion of pipeline have been formalized by proposing streamalgebras [6, 7, 10, 9]. Stream algebras formally def i ne pipelines.Each pipeline can be seen as a graph, whose vertices represent op-erators and edges represent data f l ow connections. Operators havewell-def i ned semantics, which allows us to declare pipelines for-mally and analyze them mathematically. Given a suitable, well-designed stream algebra, it is possible to build general methods foroptimizing performance, implementing feedback control, and en-abling dynamic reconf i guration.Helala et al. [14] proposed a stream algebra for visual streams.Unlike most stream algebras found within the database literature,this algebra is well-suited to deal with high-volumes of unstruc-tured data that is characteristic of visual streams [14, 13]. Theyshowed that several online computer vision methods [20, 21, 19,8, 16] can be expressed using this algebra. Here we present a f i rstimplementation of their algebra. The implementation def i nes a pro-gramming framework for building distributed pipelines for com-puter vision systems using the algebraic operators of [14, 13]. Wepresent a case study of a road boundary detection system imple-mentedusingtheframeworkproposedherein. Run-timestatistics—throughput and latency—gathered from the road boundary detec-tion system highlights the advantages of feedback control and dy-namic reconf i guration features of our framework. These resultssuggest that it is advantageous to implement the road boundary de-tection system using the proposed framework.2. RELATED WORKAs stated earlier, formal algebras for data stream processing f i rstappeared within the database literature. Demers et al. [10] devel-oped a stream algebra as a declarative language for querying eventstreams. Chkodrov et al. [9] proposed a stream algebra for map-ping relational algebra operators onto their stream counterparts.