Smart Camera (c) on board processing and wireless communication capability.

3. Smart Camera (c) • on board processing and wireless communication capability.

4. Smart Camera (c) • on board processing and wireless communication capability. Vision Sensor Network, VSN

5. Distributed Volumetric Reconstruction Shape from apparent contours Distributed Smart Cameras Camera Communication Network(CCN) optimization problem

6. Outline • Over-all Procedure • Camera-centric Distributed Algorithm • Shape From Apparent Contours • Job Distribution Scheme • Communication Optimization • Experimental Results

7. Over-all Procedure • Problem with general parallel implementation. • Camera-centric distributed algorithm. • The camera c only maintain subset Vc.

8. Camera-centric Distributed Algorithm • Compute Fvc • Terminate • Send Fvcto all the cameras in Cv • Update voxel’s level set value • Update Vc Max|Fvc| < ε Yes No

9. Camera-centric Distributed Algorithm • Compute Fvc • Terminate • Send Fvcto all the cameras in Cv • Update voxel’s level set value • Update Vc Max|Fvc| < ε Yes No

10. Camera-centric Distributed Algorithm • Compute Fvc • Terminate • Send Fvcto all the cameras in Cv • Update voxel’s level set value • Update Vc Max|Fvc| < ε Yes No • Update PRS (Primary responsibility set) • Update SRS (Secondary responsibility set)

11. Camera-centric Distributed Algorithm • Compute Fvc • Terminate • Send Fvcto all the cameras in Cv • Update voxel’s level set value • Update Vc Max|Fvc| < ε Yes No Minimum Spanning Tree (MST) message passing protocol.

12. Shape From Apparent Contours • Proposed algorithm combines in 2D active contours and variational surface reconstruction based on implicit surface deformation. • In active contour fitting,

13. Shape From Apparent Contours • Constraint with contour generators,

14. Shape From Apparent Contours • Occluding geometry relationship between normal and tangent plane (got from back-projecting) • Minimizing the function to reconstruct surface S*

15. Shape From Apparent Contours • Minimizing by gradient descent methods. • Rewrite with gradient descent flow and evolution equation (zero set) • Through some derivation, get the speed function

16. Job Distribution Scheme • Implementation: band (interval [DL, DH]) • Divide band into patch (one camera take care of one patch) PRS：Primary responsibility set SRS：Secondary responsibility set

17. Job Distribution Scheme • Implementation: band (interval [DL, DH] on each ψ(v)) • Divide band into patch (one camera take care of one patch) • Each patch contains at least all the “core voxels” and some of “free voxels” PRS：Primary responsibility set SRS：Secondary responsibility set

18. Update the PRS • Each iteration, given the new detected narrow band • Delete voxels in the watching list whose ψ(v) outside [DL,DH] else adding to boundary list. • Put voxels into PRS whose indicator value > threshold TG

19. Update the SRS • Two consideration • Put voxels to the cameras that voxels may belong to in next iteration. • Save communication • Load balance

20. Update the SRS

21. Update the SRS • Communication occur • Complexity: O(num of boundary voxels)

22. Tree message passing protocol Synchronization is implicitly controlled by the message passing.

23. Experimental Results • 23 images with simple background. • 56x120x96 grids. • 200 iterations.

24. Experimental Results • 20 images in nature indoor background. • 64x64x64 grids. • 200 iterations. • No need intensity matching(MVS).

25. Experimental Results • 20 images in nature indoor background. • 64x64x64 grids. • 200 iterations. • No need intensity matching(MVS).

26. Experimental Results • With 200 iterations, total exchanged is about 13.6MB(dinosaur).