Buildings Recognition and Camera Localization Using Image Texture Description
Download as PPT, PDF0 likes698 views
The document discusses a method for recognizing buildings and localizing cameras using image texture descriptions at the 25th International Cartographic Conference. It outlines methodologies for enhancing GIS databases with unique building information, highlighting techniques for facade recognition and camera geolocation, as well as potential applications and limitations. Key challenges include false matching during facade recognition and distance/angle constraints for accurate camera positioning.
![Buildings Recognition and Camera Localization Using Image Texture Description SULEIMAN Wassim 1 , JOLIVEAU Thierry 1 , FAVIER Eric 2 1 ISTHME-ISIG CNRS/UMR EVS, Université Jean Monnet - Saint-Etienne. 2 DIPI EA 3719 École Nationale d'Ingénieurs de Saint-Etienne [email_address] [email_address] [email_address] 25th International Cartographic Conference (Sageo) – 8 july 2011 – Palais de congrès Paris](https://image.slidesharecdn.com/sageoanglais11-110909155109-phpapp01/85/Buildings-Recognition-and-Camera-Localization-Using-Image-Texture-Description-1-320.jpg)
![Thank you For your attention Suleiman wassim [email_address]](https://image.slidesharecdn.com/sageoanglais11-110909155109-phpapp01/85/Buildings-Recognition-and-Camera-Localization-Using-Image-Texture-Description-28-320.jpg)
Buildings Recognition and Camera Localization Using Image Texture Description
- 1. Buildings Recognition and Camera Localization Using Image Texture Description SULEIMAN Wassim 1 , JOLIVEAU Thierry 1 , FAVIER Eric 2 1 ISTHME-ISIG CNRS/UMR EVS, Université Jean Monnet - Saint-Etienne. 2 DIPI EA 3719 École Nationale d'Ingénieurs de Saint-Etienne [email_address] [email_address] [email_address] 25th International Cartographic Conference (Sageo) – 8 july 2011 – Palais de congrès Paris
- 2. Objective Find a building in an image SIG 3D 3D GIS Locate the camera that took the image by using the location of the building
- 3. Methodology Enhancing the GIS database with information which can describe the building unique information quantifiable information The texture signature Texture signature
- 4. How to isolate the building facade in the image? Manual method (long) Sourimant 2009 Automatic method : (3D SIG model/2D image) registration (simple building)
- 5. Work plan GIS database enhancement with building texture information Facade recognition Camera geolocation Possible applications Limits
- 6. Work plan Enhancing GIS databases with building texture information Facade recognition Camera Geolocation Possible Applications Limits
- 7. Enhancing GIS databases with building texture information Texture analyses (SIFT) Finding the interest points with their local descriptor
- 8. Enhancing GIS databases with building texture information Finding the (x,y,z) of the interest points Homography constraints 3D GIS model
- 9. Enhancing GIS databases with building texture information The texture descriptor : list of interest points with their local descriptor and their 3D position
- 10. Work plan Enhancing GIS databases with building texture information Facade recognition Camera Geolocation Possible Applications Limits
- 11. Facade recognition False matching because of the locality of the descriptor
- 12. Facade recognition Eliminate the false matching using the homography constraints Select the best matching score between the current image and the stored descriptor in the databases
- 13. Facade recognition The facade in the 3D GIS
- 14. Work plan Enhancing the GIS database with building texture information Facade recognition Camera geolocation Possible applications Limits
- 15. Camera Geolocation Association of the interest points with the 3D position of the matched points in the GIS databases
- 16. Camera geolocation 4 points non-collinear (Yang & al. 2009) Real position Measured position
- 17. Camera Geolocation Error for distance (20-100)m and angle (0-30°) camera direction and facade normal : Position : 1 - 3 m Orientation : 5 - 10°
- 18. Error because the texture description is not an affine function 4 points non-coplanaires SOFTPOSIT (David et al. 2004) Camera Geolocation Real position Measured position
- 19. Work plan Enhancing the GIS database with building texture information Facade recognition Camera geolocation Possible applications Limits
- 20. Possible Applications Management of photos taken in urban areas Link GIS
- 21. Possible Applications Navigation systems support in an urban environment Satellites visibility Multipath
- 22. Possible applications Initial phase in the (2D/3D) registration Sourimant 2009
- 23. Work plan Enhancing the GIS database with building texture information Facade recognition Camera geolocation Possible applications Limits
- 24. Limits Angle between camera direction and facade normal has to be less than 30°
- 25. Limits Distance between camera and facade has to be less than 200 m
- 26. Limits Identical facades :
- 27. Limits Glass facade which reflects the sky and other buildings
- 28. Thank you For your attention Suleiman wassim [email_address]
Editor's Notes
- #5: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION
- #6: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION
- #7: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION
- #8: SURF is faster than SIFT descriptor. But our tests show that the SIFT descriptor gives better results when the angle of view vary between two images SIFT code of VLFEAT and SURF code from OPENSURF. http://www.vlfeat.org/~vedaldi/code/sift.html http://www.mathworks.com/matlabcentral/fileexchange/28300
- #9: SURF is faster than SIFT descriptor. But our tests show that the SIFT descriptor gives better results when the angle of view vary between two images SIFT code of VLFEAT and SURF code from OPENSURF. http://www.vlfeat.org/~vedaldi/code/sift.html http://www.mathworks.com/matlabcentral/fileexchange/28300
- #10: SURF is faster than SIFT descriptor. But our tests show that the SIFT descriptor gives better results when the angle of view vary between two images SIFT code of VLFEAT and SURF code from OPENSURF. http://www.vlfeat.org/~vedaldi/code/sift.html http://www.mathworks.com/matlabcentral/fileexchange/28300
- #11: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION
- #12: The level of the recognition based on the number of the interest points and their discriminative descriptor
- #13: The level of the recognition based on the number of the interest points and their discriminative descriptor
- #14: The level of the recognition based on the number of the interest points and their discriminative descriptor
- #15: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION
- #17: With four points (three non collinear scene points or non collinear image points). (Yang et al. 2009) is applied. It is based on the homography matrix between the front wall in the query image and the 3D model. The best result is obtained when the camera is located just in front of the building at a distance less than 100 meters. The quality falls down steeply also if the distance between the camera and the building exceeds 70 m (low resolution), and the view angle is above 30 degrees (non affine descriptor).
- #18: With four points (three non collinear scene points or non collinear image points). (Yang et al. 2009) is applied. It is based on the homography matrix between the front wall in the query image and the 3D model. The best result is obtained when the camera is located just in front of the building at a distance less than 100 meters. The quality falls down steeply also if the distance between the camera and the building exceeds 70 m (low resolution), and the view angle is above 30 degrees (non affine descriptor).
- #19: We need at least 4 non coplanar points. The SOFTPOSIT algorithm (David et al. 2004) is used which has a high efficiency If one of the discovered facades shows an angle above 40 degrees, this facade will cause more error in the computation and contribute to the drop down of accuracy in the position of the camera
- #20: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION
- #21: Approximate position and orientation of the camera are the first step of an automatic registration process Urban environment with low GPS precision (Li et al. 2006) (Bioret et al. 2009). Affine descriptor (Mikolajczyk & Schmid 2004) for the facade texture. This could greatly ameliorate the result for multiple facades detection If the method should be applied on Smartphones with continue video data, the SURF descriptor could perform better than SIFT because it is faster ( http://www.kooaba.com/ )
- #22: Approximate position and orientation of the camera are the first step of an automatic registration process Urban environment with low GPS precision (Li et al. 2006) (Bioret et al. 2009). Affine descriptor (Mikolajczyk & Schmid 2004) for the facade texture. This could greatly ameliorate the result for multiple facades detection If the method should be applied on Smartphones with continue video data, the SURF descriptor could perform better than SIFT because it is faster ( http://www.kooaba.com/ )
- #23: Approximate position and orientation of the camera are the first step of an automatic registration process Urban environment with low GPS precision (Li et al. 2006) (Bioret et al. 2009). Affine descriptor (Mikolajczyk & Schmid 2004) for the facade texture. This could greatly ameliorate the result for multiple facades detection If the method should be applied on Smartphones with continue video data, the SURF descriptor could perform better than SIFT because it is faster ( http://www.kooaba.com/ )
- #24: The results are not always good, because of the error in the 3D model, the complexity of urban environment and the collusion Sourimant et al., 2009. GPS, GIS AND VIDEO REGISTRATION FOR BUILDING RECONSTRUCTION