If the presenter shows a particular slide more than once then that slide is mapped to multiple PR images. In that case multiple video key frames are inserted for a given slide. Such an example is on the 3rd slide of the handouts shown in Figure 1. Each key frame is also linked to a CGI script that starts the video and slide replay in a web browser.
4. EXPERIMENTAL RESULTS
Our presentation database contains 343 recorded presentation sessions. The total number of screen images captured by the Presentation Recorder is 44958. All these images are timestamped and indexed with the OCR output.
The test set is composed of 53 handout documents containing 1341 slides. These were printed before various presentations, distributed to several users for note taking, and collected afterwards for scanning. The handouts were printed using various layouts in PowerPoint (e.g., 2, 3, 4, 6, and 9 slides per page) in color, grayscale, and black&white. Some handouts contain slides with dark foreground on a light background and some others contain light foreground on a dark background. The paper handouts were scanned at 200 dpi binary on a device with an automatic sheet feeder.
The first set of experiments shows the accuracy of handout detection with template matching. The results are presented in Table 1. Handout detection was applied to 53 scanned handouts and 956 scanned document images in the UW database, which is a commonly used test set for document image analysis research. 100% of the handout documents and 99.7% of the regular documents are correctly classified. The documents that are incorrectly classified as handouts contain large tables that are similar in appearance to a PowerPoint handout layout.
After handout detection, each handout are matched against the 343 recorded presentations in the database using n-grams with n=2. The accuracy of presentation level matching is 100%. Once a presentation is identified that matches to a given handout, the segmented slide images are matched against all the Presentation Recorder images captured in that session. The retrieval results are presented in Table 2. 53 slide handouts are automatically segmented to obtain 1341 scanned slides. The total number of captured presentation recorder images, i.e. 6115, is significantly larger than the number of slides as these include many screen shots, such as videos and demos, which are not actual slides. The total number of matching presentation recorder images that are marked as ground truth items is 1474, which is still larger than the number of actual slides, because some slides are shown more than once during the presentations. The total number of correctly matched presentation recorder images is 1406, yielding a 95% overall matching accuracy. As can be seen from the table, the handouts with 4- and 6-slide layouts are retrieved with a higher accuracy, 96% and 97% respectively, then the other handouts. Based on our observations, this is because of two reasons: the segmented slide regions are large enough to make an accurate match and when 4- or 6-slide-per-page layouts are used, users do not put markings in the slide region of the handout as often as they do when they use handouts with 2- or 3-slide-per-page layouts, which provides a better matching accuracy.
5. CONCLUSIONS
In this paper, we presented a new system for creating personalized e-presentation documents by using presentation handouts as templates and populating them with a user's handwritten notes and e-presentation media. Experimental results demonstrated the accuracy of techniques for creating the e-presentation document. Delivery of such personalized multimedia documents is a valuable alternative to web-based access and retrieval of e-presentation media.
6. REFERENCES
[1] R A. Steinmetz, "Media and Distance: A Learning Experience", IEEE Multimedia, pp. 8-10, 2001. [2] Jason A. Brotherton, Gregory D. Abowd , "Lessons learned from eClass: Assessing automated capture and access in the classroom", ACM Trans. on CHI, pp. 121 - 155, 2004. [3] P. Chiu, J. Boreczky, A. Girgensohn, D. Kimber, "LiteMinutes: An Internet-Based System for Multimedia Meeting Minutes," Proc. WWW10, pp.140-149, 2001. [4] P. Chiu, J. Foote, A. Girgensohn, and J. Boreczky, "Automatically linking multimedia meeting documents by image matching," ACM Hypertext, pp. 244-245, 2000. [5] D. Bargeron, J. Grudin, A. Gupta, E. Sanocki, et al, "Asynchronous Collaboration Around Multimedia Applied to On-Demand Education," Proc. HICSS, 2001. [6] Richard Anderson, Ruth Anderson, C. Hoyer, B. Simon, et al, "Lecture Presentation from the Tablet PC," Workshop on Advanced Collaborative Environments, 2003. [7] K.N. Truong and G.D. Abowd, "StuPad: Integrating student notes with class lectures," Proc. of CHI, pp.208-209, 1999. [8] B. Erol, J. J. Hull, J. Graham, and D-S. Lee "Prescient Paper: Multimedia Document Creation with Document Image Matching", IEEE ICPR Conference, 2004. [9] Adobe Acrobat SDK, http://partners.adobe.com/public/ developer/acrobat/devcenter.html [10] University of Washington English Document Image Database, info available at http://documents.cfar.umd.edu/ resources/database/ UWASH_Database
