Graph Cuts with CUDA

More documents

Recommendations

Info

Graph Cut Algorithms • Ford-Fulkerson – Find augmenting paths from source to sink – Global scope, based on search trees – Most used implementation today by Boykov et al. • Goldberg-Tarjan (push-relabel) – Considers one node at a time – Local scope, only direct neighbors matter – Inherently parallel, good fit for CUDA
Push-Relabel in a nutshell • Some definitions – Each node x: • Has excess flow u(x) and height h(x) • Outgoing edges to neighbors (x,*) with capacity c(x,*) – Node x is active: if u(x)> 0 and h(x)< HEIGHT_MAX – Active node x • can push to neighbor y: if c(x,y) > 0, h(y) = h(x) – 1 • is relabeled: if for all c(x,*) > 0, h(*) >= h(x)
Page 1 and 2: Graph Cuts with CUDA San Jose| 02/1
Page 3 and 4: Problems solvable with Graphcuts St
Page 5 and 6: Example: Binary Segmentation Proble
Page 7 and 8: Maximum Flow = Minimum Cut Source P
Page 9: Graph Cut Solution Input Result
Page 13 and 14: Relabel Pseudocode void relabel(x,
Page 15 and 16: Graph setup Excess Flow Height 3 So
Page 17 and 18: Initialized HEIGHT_MAX = 5 Source 3
Page 19 and 20: After Push Source 6/0 3/3 0/8 1/1 -
Page 21 and 22: After Relabel Source 6/0 3/3 0/6 1/
Page 23 and 24: After 3 more Iterations, Terminated
Page 25 and 26: Graph Cut and Solution 3 Source 9 5
Page 27 and 28: CUDA Implementation • Datastructu
Page 29 and 30: Relabel Data Access Patterns • Re
Page 31 and 32: Naive CUDA Implementation • Itera
Page 33 and 34: Push, improved • Idea: - Work on
Page 35 and 36: Wave Push Active Thread Push direct
Page 37 and 38: Wave Push Active Thread Push direct
Page 39 and 40: Wave Push • After tile pushing, b
Page 41 and 42: Relabel 5 0 3 9 0 0 1 7 • Compres
Page 43 and 44: Tile based push-relabel 80% 70% 60%
Page 45 and 46: Tile based push-relabel • Repeat
Page 47 and 48: Global Relabel • BFS from sink -
Page 49 and 50: Global Relabel • Initialize all p
Page 51 and 52: Results • Comparison between Boyk
Page 53 and 54: Example Application: GrabCut
Page 55: Summary • Introduction to Graph C

Graph Cuts with CUDA

Create successful ePaper yourself

Delete template?

Save as template?