1. Source: Computer Vision and Pattern Recognition Workshops (CVPRW), 2010 IEEE Computer Society Conference on Author: Paucher, R.; Turk, M.; Adviser: Chia-Nian Shyi Speaker: 趙家寬 Date: 2010/12/28 1

2. Introduction  Two categories of Augmented reality: Pose computation Object recognition  A more flexible approach is the 6 DOF pose of the sensor. 2

3. Introduction  Creating a local database for use.  Match the two images to find point correspondences. 3

4. Database building  For each image, store the pose of the camera(rotation and position) and its parameters.  Extract Speeded-Up Robust Features(SURF) features.  Use a stereo camera for the database images.  Store the 3D position of the features. 4

5. Cell phone sensors and pose estimation  The N97 phone from Nokia(include GPS, an accelerometer, a magnetometer and a rotation sensor).  The accelerometer outputs the second derivative of the position.  The three channels that represent the acceleration along the three cell phone axes.(acceleration due to the user, gravity acceleration) 5

6. Cell phone sensors and pose estimation  Obtain the tilt of the phone by two parameters of three of the cell phone rotation matrix.  The advantage of using this sensor rather than the accelerometer is that the outputs of this sensor are not sensitive to user movements. 6

7. Cell phone sensors and pose estimation  The last parameter to compute is the amount of rotation around a vertical axis.  The magnetometer(digital compass).  It outputs the angle that gives one parameter which is enough to compute the full orientation of the cell phone. 7

8. Database search restriction  Use two criteria to select only the relevant images.  For a database image to be considered, its center(3D point) has to be seen by the cell phone camera.  Because there is an uncertainty on the phone orientation estimated by the sensors, have to extend somewhat the field of view of the cell phone camera 8

9. Database search restriction 9

10.  The second criterion prevents bad image matching configurations.  The step is done using features that are rotation-invariant to some extent.  If there is too much rotation between the two images will fail. 10

11. Outlier removal process  The most common way to remove outliers between two views is using a Random Sample Consensus(RANSAC).  Fewer than 8 points are too time- consuming, even if they are more precise.  The 8-point algorithm uses a linear criterion that gives very poor results in the presence of noise. 11

12. Outlier removal process  First remove the most obvious wrong matches by Least Median of Square.  Then use the 8-point algorithm combined with RANSAC to remove the final outliers. 12

13. Pose computation  A set of presumably correct matches between the cell phone and the database images, which enables to compute the relative pose between the two images.  The reprojection error of the database 3D points in the cell phone image is the most meaningful criterion to minimize. 13

14. Pose computation  Initialization  Final pose estimation 14

15. Results and performance  All the 3D virtual objects used for the AR experiments are planar rectangles. 15

16. Results and performance  Better results when the rotation is lower(rotation is close to 0 degrees). 16

17. Conclusion  There is no location estimation from the sensors, the user has to start the application at a known location, otherwise the application has to search all the database images.  Porting the code to an architecture that would have native support for floating- point numbers would improve the frame rate drastically. 17