Log in / Register
Home arrow Computer Science arrow Virtual, Augmented and Mixed Reality
< Prev   CONTENTS   Next >

5 Experiments and Evaluation

5.1 Page Detection Evaluation

We conduct the experiment to evaluate the efficiency of our proposed method by matching local features extracted from images in the dataset to compare the accuracy and performance of the proposed process with the original method which does not filter out keypoints and the other state-of-the-art saliency detection methods. Since the proposed process is independent of the keypoint extraction and recognition algorithms, experiments are conducted to evaluate our approach using four popular local features: BRIEF [17], BRISK [18], SIFT [19], and SURF [20].

Experiment is conducted in a system using CPU Core i3 3.3 GHz (with 4GB RAM). Our dataset consists of 200 pages (with resolution 566 x 750) of reference materials for students in Computer Science, including MSDN Magazine, ACM Transaction Magazine, and IEEE Transaction Magazine. Each typical page includes three types of regions: background, text region, and image.

All local features are extracted in two scenarios: extracting all keypoints and extracting only keypoints in important regions. Image matching is then performed with each pair of images. The accuracy of matching is computed as proportion of correctly matched pairs of images over the number of image pairs. The result of this experiment is shown in Figure 7(a).

On average, the proposed method outperforms conventional methods up to 7%. Especially, when using SIFT feature, the accuracy is boosted approximately 22%. Moreover, our saliency detection module is replaced by different existing state-of-the-art methods such as BMS [12], FT [7], GC [13], HC [14], LC [15], and SR [16] to evaluate efficiency of our approach. In most cases, our process can provide better results than others. Incorporating our pre-process stage can not only preserve the robustness of conventional methods but also boost up the accuracy.

Fig. 7. Accuracy and performance of page detection of printed reference materials

In addition, the experiments also show that our method outperforms other algorithms with all common local features (c.f. Figure 7.B). On average, using SIFT, our method is 10.3 times faster than conventional method with no filtering out keypoints. Similarly, using BRIEF and SURF, our method is 11 and 15 times faster, and especially that of using BRISK features is more than 19.4 times.

Overall, our approach does not only boost up the running time up 19.4 times but also increases the accuracy of recognizing magazines to 22%. This is the crucial criteria for real-time AR system for magazines, books, and newspapers.

Found a mistake? Please highlight the word and press Shift + Enter  
< Prev   CONTENTS   Next >
Business & Finance
Computer Science
Language & Literature
Political science