Ralph Niels, Don Willems and Louis Vuurpijl Introducing of handwritten icons Introducing the NicIcon database of handwritten icons NicI Icon Ralph Niels Don Willems Louis Vuurpijl Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Overview • Introduction • Domain: crisis management • Data collection • Icon design • Method • Data • Classification experiment • Method • Results • Conclusion and URLNEW! = Introducing the NicIcon database of handwritten icons Crisis management (CM) • Scenario: tunnel disaster • Distributed computer systems to support CM • Multiple modalities: speech, gestures, and pen The car is here Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Iconic pen gestures • Faster than handwriting • Easy to learn & remember • Visual meaningful shape Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Ralph Niels Don Willems Louis Vuurpijl Database • Public databases available for several handwriting applications • Not for iconic pen gestures • Based on symbology reference by USA Homeland Security Workgroup – Used in e.g., USA, Australia, New-Zealand Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Icons Flood Electricity Accident Car Casualty Injury Fire Gas Bomb Roadblock Fire brigade Paramedics Police Person Introducing the NicIcon database of handwritten icons Data collection • • • • 35 participants Online and offline Variation in size Per person: – 22 pages – 55 instances / icon Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Database content • Online: 26,163 iconic gestures (24,144 reported in paper) Tablet: Wacom Intuos2 A4 oversize Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Database content • Offline: 770 scanned pages Scanner: HP Scanjet 7400C flatbed (@ 300dpi, 24 bit colour) Ralph Niels Don Willems Louis Vuurpijl Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Classification experiment Data sub sets • Stratified Train set (36%) Complete Testdata set set Evaluation set (26,163 (24%) icons) (40%) • Writer dependent (WD), writer independent (WI) WD Train set (36%) WD Test set (24%) WD Evaluation set (40%) WI Train set (36%) WI Test set (24%) WI Evaluation set (40%) Introducing the NicIcon database of handwritten icons Classification experiment on online data Multi classifier system • Feature sets • 28 geometric features (Willems & Vuurpijl) • 30/60 coordinates running features (Schomaker & Vuurpijl) • 1185 features (Willems, Niels, Van Gerven, Vuurpijl) NEW! • Feature classifiers • Support Vector Machine • Multi-Layered Perceptron • Template matching • Dynamic Time Warping (Niels & Vuurpijl) Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons New feature set (‘m-fs’) • Features from literature – Geometrical, temporal, pressure D. Rubine, Specifying gestures by example, Computer Graphics. 25 (4) (1991) 329–337. J. LaViola Jr. & R. Zeleznik, A practical approach for writerdependent symbol recognition using a writer-independent symbol recognizer, IEEE Transactions on pattern analysis and machine intelligence. 29 (11) (2007) 1917–1926. L. Zhang & Z. Sun, An experimental comparison of machine learning for adaptive sketch recognition, Applied Mathematics and Computation. 185 (2) (2007) 1138–1148. and many others… Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons New feature set (‘m-fs’) • Over complete icon, but also… – Mean over strokes – Std. dev. over strokes Feature definitions in technical report at website Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons New feature set (‘m-fs’) • Feature selection – 1185 features, which are the best? – Sort them on best individual performance – Add 1 by 1 to classifier • Performance maximizes at: – 545 features for WI – 660 features for WD • Selected features: – ± 1/3 full icon – ± 1/3 mean – ± 1/3 standard deviation The best features Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Ralph Niels Don Willems Louis Vuurpijl Breaking news: the best features Area Sine first / last sample Length of diagonal Vertical offsets Average centroidal radius Introducing the NicIcon database of handwritten icons Ralph Niels Don Willems Louis Vuurpijl Classification results Classifier Features WD WI SVM g-28 84.8 73.0 m-fs 99.3 96.4 af-30 98.1 90.3 af-60 98.0 98.6 g-28 84.0 78.5 m-fs 98.7 96.3 af-30 94.5 85.2 af-60 94.4 85.1 DTW - 98.4 93.6 MCS - 99.51 97.83 MLP g-28: 28 geometric features (Willems & Vuurpijl) m-fs: 1185 features NEW! (Willems, Niels, Van Gerven, Vuurpijl) af-30/af-60: 30/60 running features (Schomaker & Vuurpijl) DTW: Dynamic Time Warping (Niels & Vuurpijl) MCS: Multiple classifier system (majority voting) Introducing the NicIcon database of handwritten icons Misclassifications • Wrong box • Sloppy drawing • Retracing Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons Discussion • Results already quite good (WD: 99.5%, WI: 97.8%), but gain is still possible • Which features are important? – For different domains • Performance in interactive experiment • Offline data still open • Mapping online -> offline Ralph Niels Don Willems Louis Vuurpijl Introducing the NicIcon database of handwritten icons http://unipen.nici.ru.nl/NicIcon Freely available: • Online (Unipen) • Offline (PNG) • Technical report about new feature set Ralph Niels Don Willems Louis Vuurpijl