Parallel implementation of geodesic distance transform with application in superpixel segmentation
Download as PPTX, PDF1 like835 views
This document presents a parallel implementation of the geodesic distance transform (GDT) for superpixel segmentation, highlighting its efficiency and edge-following capabilities. By utilizing OpenMP, the proposed method achieves significant speedups on multi-core processors while maintaining effective segmentation quality. Comparisons with existing methods demonstrate that the GDT method produces more regular superpixels faster than state-of-the-art techniques.
![Parallel implementation of geodesic distance
transform with application in superpixel segmentation
Tuan Q. Pham
Canon Information Systems Research Australia (CISRA)
tuan.pham@cisra.canon.com.au
Geodesic distance
Parallel distance transform
Geodesic distance between two points = sum of
pixel costs along a minimum-cost path
Geodesic distance transform d(cost image f,
seed points) = geodesic distance from every
pixel to its nearest seed
destination
1
0.8
0.4
minimum path, cost = 1.7
straight path, cost = 11.1
Segment image into superpixels using GDT where
The algorithm can be parallelised if multiple
passes are allowed (OK since GDT is iterative)
Cost image = gradient energy + a small offset
Seed points = well-separated local gradient
minima by adaptive non-maximum suppression
Distance transform is propagated across bands in
a next iteration (may require more iterations)
8
0.6
source
Within a pass, chamfer algorithm is sequential
Image is divided into bands for parallel processing
10
Superpixel segmentation
6
seed
4
0.2
2
0
Fig. 1. Cost image (left) and its geodesic distance transform (right)
Geodesic distance transform
Chamfer distance algorithm = multiple iterations of
a forward
propagation
Fig. 4. Band-based image partitioning for parallel GDT implementation
Fig. 6. 1000 geodesic superpixels from a 2MegaPixel image (1936×1288)
OpenMP implementation
Segmentation comparison
OpenMP = an easy to use Open Multi-Processing
platform that is designed for multicore processors
and is supported by most compilers
and
a backward
propagation
Geodesic superpixel is faster & follows edges better
OpenMP parallelises loop by compiler directives
Frame 8
SLIC [1] (4.6 seconds)
Geodesic (0.64 sec)
TurboPixels [2] (207 sec)
Fig. 7. Three state-of-the-art superpixel methods on 2MP image in Fig.6
Fig. 2. One iteration of a forward pass (left) and a backward pass (right)
Geodesic Voronoi tessellation
Geodesic distance transform (GDT) produces edgefollowing Voronoi tessellation if edge is used as cost
frame 12
frame 8
frame 4
frame 1
Evaluation of parallel GDT
Best with static scheduling (where bands are
assigned to threads in a round-robin fashion)
Cost image & 4 seed points
Input image
Method
#
Time Platform
Method
Watershed 1008 3.2s C/Matlab
Number of fwd+bwd propagation iterations
increases slightly under parallel implementation
(10 iterations are often enough for segmentation)
Entropy 1000 6.5s
C
Lattice
FH
Time Platform
Method
#
Time Platform
1024 2.3s
C
Quickshift
992 13.3s
C
Geodesic 1000 0.3s
C
CVT
1000 2.7s
Matlab
C
Turbo
1067 58.1s Matlab
C
1024 1.4s
#
SLIC
990 1.2s
Fig. 8. Segmentation of 1MP image (# denotes number of superpixels returned)
Sub-second runtime on 5 MP image or smaller
Speedup of 1.3× on 2-core, 2.6× on 4-core CPU
region without a nearest seed
GDT after 10 fwd+bwd
propagation iterations
fragmentation
2
1.5
Summary
3.5
without OpenMP
static schedule
dynamic schedule
3
speedup factor
Intermediate GDT after
a first backward pass
runtime (seconds)
Intermediate GDT after
a first forward pass
1
0.5
We proposed a parallel implementation of
2.5
2
geodesic distance transform using OpenMP
1.5
1
static schedule
dynamic schedule
0.5
0
0
1000
2000
3000
4000
0
0
1000
image width (pixels)
Nearest seed label after Nearest seed label after
a first backward pass
a first forward pass
Nearest seed label
after 10 iterations
Fig. 3. Geodesic distance transform (2nd row) and tessellation (3rd row)
Runtime
2000
3000
4000
image width (pixels)
Speedup factor
Fig. 5. Runtime & speedup factor on 2.8GHz quad-core CPU with 12GB RAM
References: 1. Achanta et al., SLIC superpixels compared to state-of-the-art superpixel methods, PAMI 34(11), 2012.
Our geodesic segmentation method
produces more regular, edge-following
superpixels at orders of magnitude faster
than state-of-the-art segmentation methods.
2. Levinshtein et al., TurboPixels: Fast superpixels using geometric flows, PAMI 31(12), 2009.
Contact details: Tuan Q. Pham (tuan.pham@cisra.canon.com.au), 1 Thomas Holt drive, North Ryde, NSW 2113, Australia
Presented at Int’l Conf. on Digital Image Computing: Techniques and Applications (DICTA) Paper 5, Poster session 2 on Thursday 8th November, 2013. Hobart, Australia](https://image.slidesharecdn.com/dicta13poster-131225192641-phpapp02/85/Parallel-implementation-of-geodesic-distance-transform-with-application-in-superpixel-segmentation-1-320.jpg){kind=tracking}
Parallel implementation of geodesic distance transform with application in superpixel segmentation
- 1. Parallel implementation of geodesic distance transform with application in superpixel segmentation Tuan Q. Pham Canon Information Systems Research Australia (CISRA) tuan.pham@cisra.canon.com.au Geodesic distance Parallel distance transform Geodesic distance between two points = sum of pixel costs along a minimum-cost path Geodesic distance transform d(cost image f, seed points) = geodesic distance from every pixel to its nearest seed destination 1 0.8 0.4 minimum path, cost = 1.7 straight path, cost = 11.1 Segment image into superpixels using GDT where The algorithm can be parallelised if multiple passes are allowed (OK since GDT is iterative) Cost image = gradient energy + a small offset Seed points = well-separated local gradient minima by adaptive non-maximum suppression Distance transform is propagated across bands in a next iteration (may require more iterations) 8 0.6 source Within a pass, chamfer algorithm is sequential Image is divided into bands for parallel processing 10 Superpixel segmentation 6 seed 4 0.2 2 0 Fig. 1. Cost image (left) and its geodesic distance transform (right) Geodesic distance transform Chamfer distance algorithm = multiple iterations of a forward propagation Fig. 4. Band-based image partitioning for parallel GDT implementation Fig. 6. 1000 geodesic superpixels from a 2MegaPixel image (1936×1288) OpenMP implementation Segmentation comparison OpenMP = an easy to use Open Multi-Processing platform that is designed for multicore processors and is supported by most compilers and a backward propagation Geodesic superpixel is faster & follows edges better OpenMP parallelises loop by compiler directives Frame 8 SLIC [1] (4.6 seconds) Geodesic (0.64 sec) TurboPixels [2] (207 sec) Fig. 7. Three state-of-the-art superpixel methods on 2MP image in Fig.6 Fig. 2. One iteration of a forward pass (left) and a backward pass (right) Geodesic Voronoi tessellation Geodesic distance transform (GDT) produces edgefollowing Voronoi tessellation if edge is used as cost frame 12 frame 8 frame 4 frame 1 Evaluation of parallel GDT Best with static scheduling (where bands are assigned to threads in a round-robin fashion) Cost image & 4 seed points Input image Method # Time Platform Method Watershed 1008 3.2s C/Matlab Number of fwd+bwd propagation iterations increases slightly under parallel implementation (10 iterations are often enough for segmentation) Entropy 1000 6.5s C Lattice FH Time Platform Method # Time Platform 1024 2.3s C Quickshift 992 13.3s C Geodesic 1000 0.3s C CVT 1000 2.7s Matlab C Turbo 1067 58.1s Matlab C 1024 1.4s # SLIC 990 1.2s Fig. 8. Segmentation of 1MP image (# denotes number of superpixels returned) Sub-second runtime on 5 MP image or smaller Speedup of 1.3× on 2-core, 2.6× on 4-core CPU region without a nearest seed GDT after 10 fwd+bwd propagation iterations fragmentation 2 1.5 Summary 3.5 without OpenMP static schedule dynamic schedule 3 speedup factor Intermediate GDT after a first backward pass runtime (seconds) Intermediate GDT after a first forward pass 1 0.5 We proposed a parallel implementation of 2.5 2 geodesic distance transform using OpenMP 1.5 1 static schedule dynamic schedule 0.5 0 0 1000 2000 3000 4000 0 0 1000 image width (pixels) Nearest seed label after Nearest seed label after a first backward pass a first forward pass Nearest seed label after 10 iterations Fig. 3. Geodesic distance transform (2nd row) and tessellation (3rd row) Runtime 2000 3000 4000 image width (pixels) Speedup factor Fig. 5. Runtime & speedup factor on 2.8GHz quad-core CPU with 12GB RAM References: 1. Achanta et al., SLIC superpixels compared to state-of-the-art superpixel methods, PAMI 34(11), 2012. Our geodesic segmentation method produces more regular, edge-following superpixels at orders of magnitude faster than state-of-the-art segmentation methods. 2. Levinshtein et al., TurboPixels: Fast superpixels using geometric flows, PAMI 31(12), 2009. Contact details: Tuan Q. Pham (tuan.pham@cisra.canon.com.au), 1 Thomas Holt drive, North Ryde, NSW 2113, Australia Presented at Int’l Conf. on Digital Image Computing: Techniques and Applications (DICTA) Paper 5, Poster session 2 on Thursday 8th November, 2013. Hobart, Australia