Method and apparatus for removing outliers from a main view of a scene during 3D scene reconstruction
View Patent ↗A method and an apparatus for removing outliers from a main view of a scene during 3D reconstruction of a scene from multiple views of the scene. A 3D projection unit projects a 3D point of a pixel of the main view into neighboring views. A comparator then compares the distance of each of the projected 3D points in the neighboring views to the 3D point of the main view with a defined distance threshold. Based on the comparison a flagging unit assigns flags to the pixel in the main view. Finally, depending on values of the flags a rejecting unit rejects the pixel in the main view as an outlier.
1. A method for removing outliers from a main view of a scene during 3D reconstruction of a scene from multiple views of the scene, the method comprising:
projecting a 3D point of a pixel of the main view into neighboring views;
comparing the distance of each of the projected 3D points in the neighboring views to the 3D point of the main view with a defined distance threshold;
assigning flags to the pixel in the main view based on this comparison; and
rejecting the pixel in the main view as an outlier depending on values of the flags.
2. The method according to claim 1 , wherein:
a flag with a first value is assigned when the 3D point in the main view and a projected 3D point in a neighboring view are in agreement;
a flag with a second value is assigned when a conflict exists between the 3D point in the main view and a projected 3D point in a neighboring view; and
a flag with a third value is assigned when a result of the comparison is undecided for a projected 3D point in a neighboring view.
3. The method according to claim 2 , wherein a 3D point in a neighboring view is in agreement with the 3D point of the main view if their distance is below the defined distance threshold.
4. The method according to claim 2 , wherein the flag with the third value is assigned when:
no depth information is available for a pixel in a neighboring view hit by projecting the 3D point of the pixel of the main view;
projecting the 3D point of the pixel of the main view into a neighboring view does not hit the neighboring view; or
a distance larger than the defined distance threshold might be due to an occlusion.
5. The method according to claim 2 , wherein the flag with the third value is changed to a flag with the second value when the distance between the 3D point of a neighboring view and the 3D point of the main view is above the defined distance threshold but below a sum of depth errors of the pixel in the main view and the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
6. The method according to claim 2 , wherein the pixel in the main view is rejected as an outlier if the number of flags with the second value is above a conflict threshold or if the number of flags with the first value is below an agreement threshold.
7. The method according to claim 2 , further comprising:
performing an additional consistency check when a flag with the first value is assigned to the pixel; and
changing the flag with the first value to a flag with the second value when the additional consistency check fails.
8. The method according to claim 7 , wherein the additional consistency check compares a property of the pixel in the main view and a property of the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
9. The method according to claim 8 , wherein the additional property of a pixel is one of an intensity, a color, an intensity averaged for a window around the pixel, and a color averaged for a window around the pixel.
10. The method according to claim 7 , wherein the additional consistency check uses a normalized cross-correlation between a window centered at the pixel in the main view and a window centered at the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
11. The method according to claim 1 , wherein in case of a sequence of views:
the sequence is divided into distinct subsets, wherein for each subset one view is selected as a main view, whereas the remaining views are selected as neighboring views; or
a view of the sequence serves as a main view, but also as a neighboring view for another view of the sequence.
12. A non-transitory computer readable storage medium having stored therein instructions enabling removing outliers from a main view of a scene during 3D reconstruction of a scene from multiple views of the scene, which, when executed by a computer, cause the computer to:
project a 3D point of a pixel of the main view into neighboring views;
compare the distance of each of the projected 3D points in the neighboring views to the 3D point of the main view with a defined distance threshold;
assign flags to the pixel in the main view based on the comparison; and
reject the pixel in the main view as an outlier depending on values of the flags.
13. The non-transitory computer readable storage medium according to claim 12 , wherein the instructions cause the computer to:
assign a flag with a first value when the 3D point in the main view and a projected 3D point in a neighboring view are in agreement;
assign a flag with a second value when a conflict exists between the 3D point in the main view and a projected 3D point in a neighboring view; and
assign a flag with a third value when a result of the comparison is undecided for a projected 3D point in a neighboring view.
14. The non-transitory computer readable storage medium according to claim 13 , wherein a 3D point in a neighboring view is in agreement with the 3D point of the main view if their distance is below the defined distance threshold.
15. The non-transitory computer readable storage medium according to claim 13 , wherein the instructions cause the computer to assign the flag with the third value when:
no depth information is available for a pixel in a neighboring view hit by projecting the 3D point of the pixel of the main view;
projecting the 3D point of the pixel of the main view into a neighboring view does not hit the neighboring view; or
a distance larger than the defined distance threshold might be due to an occlusion.
16. The non-transitory computer readable storage medium according to claim 13 , wherein the instructions cause the computer to change the flag with the third value to a flag with the second value when the distance between the 3D point of a neighboring view and the 3D point of the main view is above the defined distance threshold but below a sum of depth errors of the pixel in the main view and the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
17. The non-transitory computer readable storage medium according to claim 13 , wherein the instructions cause the computer to reject the pixel in the main view as an outlier if the number of flags with the second value is above a conflict threshold or if the number of flags with the first value is below an agreement threshold.
18. The non-transitory computer readable storage medium according to claim 13 , wherein the instructions further cause the computer to:
perform an additional consistency check when a flag with the first value is assigned to the pixel; and
change the flag with the first value to a flag with the second value when the additional consistency check fails.
19. The non-transitory computer readable storage medium according to claim 18 , wherein the additional consistency check compares a property of the pixel in the main view and a property of the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
20. The non-transitory computer readable storage medium according to claim 19 , wherein the additional property of a pixel is one of an intensity, a color, an intensity averaged for a window around the pixel, and a color averaged for a window around the pixel.
21. The non-transitory computer readable storage medium according to claim 18 , wherein the additional consistency check uses a normalized cross-correlation between a window centered at the pixel in the main view and a window centered at the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
22. The non-transitory computer readable storage medium according to claim 12 , wherein in case of a sequence of views:
the sequence is divided into distinct subsets, wherein for each subset one view is selected as a main view, whereas the remaining views are selected as neighboring views; or
a view of the sequence serves as a main view, but also as a neighboring view for another view of the sequence.
23. An apparatus configured to remove outliers from a main view of a scene during 3D reconstruction of a scene from multiple views of the scene, the apparatus comprising:
a 3D projection unit configured to project a 3D point of a pixel of the main view into neighboring views;
a comparator configured to compare the distance of each of the projected 3D points in the neighboring views to the 3D point of the main view with a defined distance threshold;
a flagging unit configured to assign flags to the pixel in the main view based on the comparison; and
a rejecting unit configured to reject the pixel in the main view as an outlier depending on values of the flags.
24. The apparatus according to claim 23 , wherein the flagging unit is configured to:
assign a flag with a first value when the 3D point in the main view and a projected 3D point in a neighboring view are in agreement;
assign a flag with a second value when a conflict exists between the 3D point in the main view and a projected 3D point in a neighboring view; and
assign a flag with a third value when a result of the comparison is undecided for a projected 3D point in a neighboring view.
25. The apparatus according to claim 24 , wherein a 3D point in a neighboring view is in agreement with the 3D point of the main view if their distance is below the defined distance threshold.
26. The apparatus according to claim 24 , wherein the flagging unit is configured to assign the flag with the third value when:
no depth information is available for a pixel in a neighboring view hit by projecting the 3D point of the pixel of the main view;
projecting the 3D point of the pixel of the main view into a neighboring view does not hit the neighboring view; or
a distance larger than the defined distance threshold might be due to an occlusion.
27. The apparatus according to claim 24 , wherein the flagging unit is configured to change the flag with the third value to a flag with the second value when the distance between the 3D point of a neighboring view and the 3D point of the main view is above the defined distance threshold but below a sum of depth errors of the pixel in the main view and the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
28. The apparatus according to claim 24 , wherein the rejecting unit is configured to reject the pixel in the main view as an outlier if the number of flags with the second value is above a conflict threshold or if the number of flags with the first value is below an agreement threshold.
29. The apparatus according to claim 24 , wherein the flagging unit is further configured to:
perform an additional consistency check when a flag with the first value is assigned to the pixel; and
change the flag with the first value to a flag with the second value when the additional consistency check fails.
30. The apparatus according to claim 29 , wherein the additional consistency check compares a property of the pixel in the main view and a property of the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
31. The apparatus according to claim 30 , wherein the additional property of a pixel is one of an intensity, a color, an intensity averaged for a window around the pixel, and a color averaged for a window around the pixel.
32. The apparatus according to claim 29 , wherein the additional consistency check uses a normalized cross-correlation between a window centered at the pixel in the main view and a window centered at the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
33. The apparatus according to claim 23 , wherein in case of a sequence of views:
the sequence is divided into distinct subsets, wherein for each subset one view is selected as a main view, whereas the remaining views are selected as neighboring views; or
a view of the sequence serves as a main view, but also as a neighboring view for another view of the sequence.
34. An apparatus configured to remove outliers from a main view of a scene during 3D reconstruction of a scene from multiple views of the scene, the apparatus comprising a processing device and a memory device having stored therein instructions, which, when executed by the processing device, cause the apparatus to:
project a 3D point of a pixel of the main view into neighboring views;
compare the distance of each of the projected 3D points in the neighboring views to the 3D point of the main view with a defined distance threshold;
assign flags to the pixel in the main view based on the comparison; and
reject the pixel in the main view as an outlier in depending on values of the flags.
35. The apparatus according to claim 34 , wherein the instructions cause the processing device to:
assign a flag with a first value when the 3D point in the main view and a projected 3D point in a neighboring view are in agreement;
assign a flag with a second value when a conflict exists between the 3D point in the main view and a projected 3D point in a neighboring view; and
assign a flag with a third value when a result of the comparison is undecided for a projected 3D point in a neighboring view.
36. The apparatus according to claim 35 , wherein a 3D point in a neighboring view is in agreement with the 3D point of the main view if their distance is below the defined distance threshold.
37. The apparatus according to claim 35 , wherein the instructions cause the processing device to assign the flag with the third value when:
no depth information is available for a pixel in a neighboring view hit by projecting the 3D point of the pixel of the main view;
projecting the 3D point of the pixel of the main view into a neighboring view does not hit the neighboring view; or
a distance larger than the defined distance threshold might be due to an occlusion.
38. The apparatus according to claim 35 , wherein the instructions cause the processing device to change the flag with the third value to a flag with the second value when the distance between the 3D point of a neighboring view and the 3D point of the main view is above the defined distance threshold but below a sum of depth errors of the pixel in the main view and the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
39. The apparatus according to claim 35 , wherein the instructions cause the processing device to reject the pixel in the main view as an outlier if the number of flags with the second value is above a conflict threshold or if the number of flags with the first value is below an agreement threshold.
40. The apparatus according to claim 35 , wherein the instructions further cause the processing device to:
perform an additional consistency check when a flag with the first value is assigned to the pixel; and
change the flag with the first value to a flag with the second value when the additional consistency check fails.
41. The apparatus according to claim 40 , wherein the additional consistency check compares a property of the pixel in the main view and a property of the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
42. The apparatus according to claim 41 , wherein the additional property of a pixel is one of an intensity, a color, an intensity averaged for a window around the pixel, and a color averaged for a window around the pixel.
43. The apparatus according to claim 40 , wherein the additional consistency check uses a normalized cross-correlation between a window centered at the pixel in the main view and a window centered at the pixel in the neighboring view hit by projecting the 3D point of the pixel of the main view.
44. The apparatus according to claim 34 , wherein in case of a sequence of views:
the sequence is divided into distinct subsets, wherein for each subset one view is selected as a main view, whereas the remaining views are selected as neighboring views; or
a view of the sequence serves as a main view, but also as a neighboring view for another view of the sequence.