SLIDE 1
An Efficient Algorithm for Feature-based 3D Point Cloud - - PowerPoint PPT Presentation
An Efficient Algorithm for Feature-based 3D Point Cloud - - PowerPoint PPT Presentation
Zili Yi Co-authors: Yang Li, Minglun Gong Memorial University of Newfoundland 2016-12-12 An Efficient Algorithm for Feature-based 3D Point Cloud Correspondence Search Outline Introduction Background Method Results
SLIDE 2
SLIDE 3
Introduction – Point Cloud
3D point cloud
A versatile representation of 3D shapes
Sources
laser scanners estimated from stereo matching sampled from 3D surface models.
SLIDE 4
Introduction - 3D Point Cloud Correspondence
3D Point Cloud Correspondence
to find the matching between
the two sets of points
Application
building statistical shape models smoothly interpolating key
Frames in cartoon animations
morphing between shapes of disparate
- bjects
recognizing/classifying 3D objects 3D Point Cloud Correspondence
SLIDE 5
Introduction – previous methods
Methods for Point Cloud Correspondence
Rigid (transformation to correspondence)
ICP
, BMP
Non-rigid (Pointwise/articulated correspondence)
Feature-based Non-rigid BMP/ICP
Feature-based Correspondence
Matching between features rather than raw points
Limitation of traditional feature-based algorithm
Computation of features at multiple levels Inefficient when introducing smoothness term
SLIDE 6
Related topics
Nearest Neighbor Search
Space partitioning trees: K-d tree,
VP-tree
PatchMatch (for 2D image correspondence search)
Combine randomized search and belief propagation
3D Point Cloud Features
Unique Shape Context Point Feature Point Feature Histogram
SLIDE 7
Related topics
Swarm Intelligence
Artificial Bee Colony (ABC) Search optimal solutions
Three types of bees in a Colony
Scout
Search food sources randomly
Employed bee
Search food sources among neighbors
Onlooker
Search food sources from other colonies
SLIDE 8
Method Framework
Objective Function Optimization Correspondence Result Preparation 3D Feature Extraction Colorization K-Neighbor Caching Random Initialization Source Target ABC-based Search Converge Source Target
SLIDE 9
Method
Objective
S: Source, T: Target P ∈ 𝑇, M(P) is the matching point of P
α is the balance coefficient
Geometric term Smooth term
SLIDE 10
Effect of Smoothness Term
Force neighbors corresponding to neighbors
SLIDE 11
Optimization Process
Searching scheme (iteratively and pointwisely)
SLIDE 12
Optimization Process
SLIDE 13
Optimization Technique
Variable harmonic vs fixed harmonic during optimization
Source Target Source α=0 α varying 0-0.95 α=0.95 (mesh) (mesh) (point cloud) result I result II result III
SLIDE 14
Qualitative Comparison
SLIDE 15
Efficiency
An order of magnitude faster than brute-force search
SLIDE 16
Conclusion
Strength
Efficient Accurate Noise-robust
Limitation
Poor matching with giant shape difference Correspondence results on the tiger-horse dataset under different noise levels. In each test, both the target (left) and the source (right) point clouds are corrupted using Gaussian noise with standard
variance of σ, where D is the diagonal length of the input point cloud.
SLIDE 17
Conclusion
Strength
Efficient Accurate Noise-robust
Limitation
Poor matching between shapes with giant difference Correspondence results on the tiger-horse dataset under different noise levels. In each test, both the target (left) and the source (right) point clouds are corrupted using Gaussian noise with standard
variance of σ, where D is the diagonal length of the input point cloud.
SLIDE 18