Presentation at ISSDQ 2015 by Guillaume Touya

Transcript

1. COMPARING IMAGE-BASED METHODS FOR ASSESSING VISUAL CLUTTER IN GENERALIZED MAPS G. Touya – B. Decherf – M. Lalanne – M. Dumont COGIT team IGN France ISSDQ 2015 – Geospatial week
2. Evaluation of map generalization Accept or cancel Initial data Select situation and algorithms Apply selected process evaluation required Map output
3. Evaluation of map generalization • User map requirements • Cartographic rules Generalization constraints
4. Evaluation of map generalization • User map requirements • Cartographic rules Generalization constraints « Building area > 0.4 map mm² » « Building granularity > 0.1 map mm » « Building/road distance > 0.1 map mm » « Building alignments should be preserved »
5. Evaluation of map generalization Constraints number 80 60 40 20 0 1 2 3 4 5 6 7 8 satisfaction scale
6. Evaluation of map generalization Constraints number 80 60 40 20 0 1 2 3 4 5 6 7 8 Global evaluation is complex! satisfaction scale
7. [Rylov & Reimer 2014] Using image-based evaluation
8. Clutter in computer vision Excessive and/or disorganized information
9. Edge density clutter
10. [Jégou & Deblonde 2012] Quad-tree based clutter
11. [Jégou & Deblonde 2012] Quad-tree based clutter
12. [Rosenholtz et al 2007] Subband entropy clutter Similar to jpeg compression
13. [Jégou & Deblonde 2012] Subband entropy clutter
14. [Bravo & Farid 2008] Segmentation based clutter
15. Tested maps Initial data 1:15k symbols Initial data 1:50k symbols Generalized data 1:50k symbols
16. Tested maps Before and after, manually generalized
17. [Touya & Duchêne 2011] Tested maps Before and after, automatically generalized
18. Experiments
19. Identifying Too Cluttered Areas Initial map quad tree edges segmentation
20. Preserving the global amount of information 1:25k 1:100k 1:250k 1:1000k Edge density Subband entropy 2000 1500 1000 500 0 10.5 10 9.5 9 8.5 8 0.15 0.1 0.05 0 Quad tree Segmentation 4000 3000 2000 1000 0
21. Preserving the global amount of information 1:25k 1:100k 1:250k 1:1000k 1500 Edge density Subband entropy 15 1000 10 500 5 0 0 0.2 0.15 0.1 0.05 0 Quad tree Segmentation 6000 4000 2000 0
22. Preserving the global amount of information 1:25k 1:100k Edge density 2000 1500 15 1:250k Subband entropy 10 1000 5 500 0 0 1:25k 1:100k 1:250k 1:25k 1:100k 1:250k Quad tree Segmentation 4000 0,1 0,08 0,06 0,04 0,02 0 3000 2000 1000 0 1:25k 1:100k 1:250k 1:25k 1:100k 1:250k
23. Handling of occlusions and overlaps Clutter measure Edge density Subband entropy Quad tree Segmentation Before generalization After generalization 6664 9268 4.1 4.39 0.042 0.062 398 389
24. Handling of occlusions and overlaps Clutter measure Edge density Subband entropy Quad tree Segmentation Before generalization After generalization 6664 9268 4.1 4.39 0.042 0.062 398 389
25. Handling of occlusions and overlaps Initial map Generalized map
26. Background and foreground Edge density Subband entropy Quad tree Segmentation initial map 1432 4.91 0.058 1076 no background 2427 2.39 0.028 431 transparency 1543 4.17 0.028 633 paler shades 542 4.95 0.028 927
27. Background and foreground Edge density Subband entropy Quad tree Segmentation initial map 1432 4.91 0.058 1076 no background 2427 2.39 0.028 431 transparency 1543 4.17 0.028 633 paler shades 542 4.95 0.028 927
28. Background and foreground Edge density Subband entropy Quad tree Segmentation initial map 1432 4.91 0.058 1076 no background 2427 2.39 0.028 431 transparency 1543 4.17 0.028 633 paler shades 542 4.95 0.028 927
29. [Olsson et al 2011] Background and foreground Including contour lines Excluding contour lines Legible cell Cluttered cell
30. Reduction of blank space Clutter measure Before generalization After generalization Edge density 429 1459 Subband entropy 1.73 1.75 Quad tree 0.032 0.035 Segmentation 1630 1765
31. Reduction of blank space Clutter measure Before generalization After generalization Edge density 429 1459 Subband entropy 1.73 1.75 Quad tree 0.032 0.035 Segmentation 1630 1765
32. Conclusion • Image-based clutter measures can be useful for generalization evaluation • Clutter measures vary differently with generalization
33. What’s next? • Test other methods (feature congestion, color clustering, crowding model…) • Compare with vector-based methods • Combine several complementary methods • Compare to generalization evaluation methods
34. Any Question? G. Touya – B. Decherf – M. Lalanne – M. Dumont COGIT team IGN France ISSDQ 2015 – Geospatial week
35. Implementation detail • All is available in open source Java platform GeOxygene • Use of OpenIMAJ library [Hare et al 2011] • Test images can be made available on demand for comparisons