1 / 27

Segmentation with raster sliding local window

Segmentation with raster sliding local window. Nelson Chang Institute of Electronics, National Chiao Tung University. Outline. Stereo Matching with Segment Traditional Watershed-based Segmentation Downfall of tradition segmentation methods Proposed Raster Sliding Local Window Method

Télécharger la présentation

Segmentation with raster sliding local window

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Segmentation with raster sliding local window Nelson Chang Institute of Electronics, National Chiao Tung University

  2. Outline • Stereo Matching with Segment • Traditional Watershed-based Segmentation • Downfall of tradition segmentation methods • Proposed Raster Sliding Local Window Method • Result • Future work

  3. Census Diffusion WITHOUT Segmentation

  4. Census Diffusion WITH segmentation

  5. Mean Shift vs. Watershed Watershed Mean Shift

  6. Traditional Watershed-based Segmentation • Steps • Generate Gradient Map • Reconstruction/Global Denoise • Watershed flooding or Rain dropping • Region Merging

  7. What makes an algorithm efficient to implement? state 0 state 0 state 3 • Regular • Simple control • High data reusability • Parallel • More scalable • Small storage • Lower cost • More fast memory usage state 1 state 1 state 2 state 2 state 4 Simple FSM Data reuse!!! PE PE PE PE PE PE 4x speed up!! PE PE Slow SDRAM Temporary Data Temporary Data Fast SRAM

  8. Traditional Watershed and Topoggan Irregular jumpy access pattern HW: hard to reuse SW: low cache hit • Watershed Flooding • Bottom-up • Finds the minimum first

  9. Traditional Watershed and Topoggan Irregular access start point HW: less chance to reuse SW: more cache miss More regular local access pattern than watershed • Topoggan Rain Dropping • Top-down • Finds the maximum peak first • Drop rain • Back trace the path Back trace water flow path (tree traversal) High data dependency HW: less parallelism available SW: not much impact

  10. Our Goal • Regular process and access pattern • Avoid tree traversal/tree traversal • Minimize iteration count

  11. Overview (1/3) • Flowchart

  12. Gradient Generation (0th Iteration) • Color space • YUV • Select the maximal gradient among the three components • Gradient - Mathematical morphology • ABS(Dilation – Erosion) • Dilation : Max in 3x3 square window • Erosion : min in 3x3 square window 16 9 8 7 10 10 Gradient = 16-5 = 11 6 10 5

  13. Local approach

  14. Local Denoise (1st Iteration) Neighbor of x  N(x) = {i, j, k, l, m, n, o, p} Gradient of x  G(x) Gradient after local denoise G’ For each N(x), If | G(x) - G(N(x)) | < TH, G’(N(x) = G(x) Else G’(N(x)) = G(x) G G’ i j k 19 24 82 19 30 82 TH=10 l x m 37 30 38 30 30 30 n o p 10 13 12 10 13 12

  15. An Example of Local Denoise • G’ (Gradient after local denoise) of an overlapping pixel in two neighboring windows may be different 19 24 82 42 TH=10 G 37 30 38 10 10 13 12 29 19 30 82 24 82 38 30 30 30 G’ 38 38 10 G’ 10 13 12 10 13 38

  16. Adaptive Bi-threshold • TH for the next window changes • Current window is a plateau/basin • Next TH = TH_max • Else • Next TH = TH_min Plateau/ Basin Use TH_max Current Next Possible Edge/Hill Use TH_min

  17. Basin/Plateau Detection (1st Iteration) Neighbor of x  N(x) = {i, j, k, l, m, n, o, p} Gradient of x  G(x) Gradient after local denoise G’ Segment (Region) Label of x  L(x) For all N(x), If G’(N(x)) – G’(x) ≥ 0, L(x) = all L(N(x)) Else L(x) = no label G’ L L 20 45 20 62 20 31 32 20 20 Basin No original label on N(x) One N(x) already has label

  18. Set label alias = Segment Label Aliasing (1st Iteration) • For a basin/plateau L Label Alias Table = = = More than one N(x) already having different label = update = = = =

  19. Label Upward Diffusion (2nd~Nth Iteration) Neighbor of x  N(x) = {i, j, k, l, m, n, o, p} Gradient of x  G(x) Segment (Region) Label of x  L(x) For each N(x), If L(N(x)) = no label AND G(N(x)) – G(x) ≥ 0, L(N(x))=L(x) Else L(N(x)) = no label This can be done for more than 1 iterations G L 19 24 82 37 30 38 10 13 12

  20. Label Replacement (N+1 Iteration) • Update each pixel’s label • Use Label Alias Table • Similar to region (segment) merging Label Alias Table = = = = = = = =

  21. Segmentation with Raster Sliding Local Window Watershed Raster Sliding Window

  22. Results

  23. Ranking • CDS stereo matching performance comparison on Middleburry website

  24. Future work • Complexity analysis • Algorithmic Complexity • Operation Complexity • SW execution • Storage Requirement • Hardware implementation • ISCAS’09 • Refine performance • Boundary improvement • Label alias issue

More Related