310 likes | 317 Views
AutoLabel enables the automatic labeling of places by matching in-store text with website content, utilizing Wi-Fi AP vectors for accurate localization.
E N D
AutoLabel: Labeling Places from Pictures and Websites Rufeng Meng, Sheng Shen, Romit Roy Choudhury, Srihari Nelakuditi
Localization Coordinates Names
AP Name Mining AP NameMining Best Buyon Street Macy’sin Mall
Mapping AP Vector to Store Name OCR AutoLabelMatching In-store Text with Website Text Matching In-store Textwith Website Text Text From Candidate Webpages Best Buy Whole Foods Panera Starbucks
Candidate Store Names CrowdSourcing m n
CrowdSourcing m n Candidate Store Names
CrowdSourcing OCRed In-store Text 1 Text-based Matching 1 Candidate Store Names
Text Extraction Above Eye-level • Store Text • Text OCRed from In-store Pictures • Web Text • Candidate Store Names • Meta Keywords in Web Pages • Text in Menu/Category Items Eye-level Below Eye-level
Noun/Proper Name Extraction Filter Weight Assignment Bag of Words Similarity Calculation Labeling Noun/Proper Name Extraction Filter Weight Assignment Bag of Words • Weight assignment: TF-IDF Text Matching Store Text Web Text
<AP6, AP7, AP8> <AP1, AP2, AP5> <AP2, AP3, AP4> Cluster (Best Buy) (Dollar Tree) (Starbucks) (Panera Bread) Webs of Candidate Stores AutoLabel In-store Textvs. Website Text
Simultaneous Clustering and Labeling • Two images put in different clusters if no common AP • Generate potential clusterings of images • Similarity of clustering = sum of similarities of its clusters • Pick the clustering with highest similarity • Label the AP vectors as per the best clustering
Data Collection • 40 Stores • 18 from shopping mall @Champaign, IL • 10 from one street area @Champaign, IL • 6 from other places @Champaign, IL • 6 from shopping mall @San Jose, CA • Number of Pics: • 6 ~ 217 pics/store(with readable text)
Text Matching without Store Names Accuracy: 80%
Text Matching with Store Names Accuracy: 87%
Text Matching without Store Names (Mall) 16/18 Correctly Matched
Text Matching with Store Names (Mall) 17/18 Correctly Matched
Text Matching without Store Names (Street) 10/10 Correctly Matched
Accuracy with Varying #Stores/Area Mean Accuracy: • 5-store: 100% • 20-store: 90%
Performance of Place Recognition with AP Vector– Store Name Mapping
Related Work • Visual business recognition • Image based matching of the exterior of store • Human assisted positioning using textual signs • Requires user intervention to localize self • Correlate crowdsourced pictures with social network posts • Stronger correlation between store and its website
Limitations & Future Work • WiFi AP tagged in-store pictures • Reduce uncertainty even with a few pictures • Common decorations besides text • Image clustering • Spatial relationship between stores • Color/Theme + AP name mining • Beyond semantic localization • Correlate website and in-store shopping behavior