%0 Conference Proceedings %F ng:safe07 %A Ng, Tian-Tsong %A Chang, Shih-Fu %A Tsui, Mao-Pei %T Lessons Learned from Online Classification of Photo-Realistic Computer Graphics and Photographs %B IEEE Workshop on Signal Processing Applications for Public Security and Forensics (SAFE) %X We presented a set of physics motivated features for classifying photographic and computer graphic images in our previous work [1]. We also deployed an online demo system for distinguishing photographic and computer graphic images in October 2005 [2], which features our geometry classifier, together with the wavelet classifier, and the cartoon classifier. On the first anniversary of its launch, we have received 1582 submitted images, through which we perform an analysis on the user behavior, the image set characteristics, and the classifier performance. We observe that online users do not provide clear judgments for about 80% of the submitted images, confirming the challenge in distinguishing photo-realistic computer graphics images from natural photographs. We also found the accuracy of our classifiers over the online submission set is consistent with that computed over an offline data set. Finally, in order to improve the online computational speed of our classifier, we perform feature selection and reduction, cutting the response time from 152 seconds to 24 seconds per image, while keeping the accuracy almost unchanged. %U http://www.ee.columbia.edu/dvmm/publications/07/ieee-safe-07-submit.pdf %8 April %D 2007