Transcript MeshValmet: Validation Metric for Triangle Meshes

Surface Comparison and
Validation Metric
Christine Xu
University of North Carolina at
Chapel Hill
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
1
Motivation
• Segmentation comparison
• Shape Statistics: compare two group means
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
2
Volume Overlap
• Dice coefficient
– For segmentation X and Y, the coefficient may be
defined as shared information (intersection) over
the average of the X and Y volumes:
• What is the range of Dice coefficient?
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
3
Volume Overlap Measurement
• Pros
• straightforward, easy to understand and compute
• Cons
– Value depends on the amount of boundary exposed
– Missing important
local shape information
Scenario 1
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
Scenario 42
Boundary Distance Measurement
• Between what?
– Ideally, corresponding boundary points
– Normally, closest points
• Interpoint distances for two sets, S and
– Thus union over both sets: d(S,
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
) and d(
, S)
5
Boundary Distance Measurement
• Pros:
– detailed, capture all local shape information
• Cons:
– Too much information: a very long distance vector
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
6
Reportable Values
• Histogram
– The distribution of surface-to-surface distances
• Mean
• Standard deviation
• Quantiles
– Divide ordered data into q essentially equal-sized
subsets, the kth q-quantile is the value x such that:
– The 10-quantiles are called deciles
– The 100-quantiles are called percentiles
• The nth percentile is the smallest number that is greater than
n% of the whole dataset
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
7
Reportable Quantiles
• Hausdorff distance: 100% quantile
– Con: reflects really weird points
• 95th percentile
– More or less what we wanted Hausdorff to tell us
• Quartiles
– 75th percentile is conservative
– 50th percentile: the median
• Often more representative central value than the mean
• More robust than the mean
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
8
MeshValmet
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
9
MeshValmet
• A tool that measures surface-to-surface distance
between two triangle meshes using user-specified
uniform sampling
– Finer sampling: more accuracy
– Sparser sampling: faster, rough feeling of error distribution
•
•
•
•
Can be run in GUI mode or batch mode
Provide histogram and statistical information
3D visualization of surface distances
Based on the code and paper of: MESH: Measuring
Errors between Surfaces using the Hausdorff distance
(ICME 2002) by Nicolas Aspert.
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
10
MeshValmet
• Documentation (pdf):
– Google ‘MeshValmet’ or:
– http://www.cs.unc.edu/~xushun/MeshValmet3.0.pdf
• Windows and Linux executables:
– Windows:
http://www.cs.unc.edu/~xushun/MeshValmet-3.0Windows.zip
– Linux:http://www.cs.unc.edu/~xushun/MeshValmet3.0-Linux.tar.gz
• Any questions or bugs, please report to me at:
[email protected].
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
11
Usage of MeshValmet
• Step 1. Load in two triangle meshes. Key ‘t’ and ‘j’ to switch
between the ‘trackerball’ and ‘joystick’ mode in two display
windows.
• Step 2. Specify sampling step and minimum sampling frequency
used for sub-sampling.
• Step 3. Choose A->B or B-> or both. This is due to the asymmetric
property of Hausdorff distance.
• Step 4. Choose the number of bins of the resulting histogram.
• Step 5. Choose either signed distance error or absolute distance
error. Positive: outward distance; negative: inward distance
• Step 6. Click button ‘Compute Error’
• Step 7. (optional) Change settings in the ‘Colormap Control’ panel
• Step 8. (optional) Change settings in the ‘Histogram Control’ panel
• Step 9. View the statistical results of the comparison.
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
12
Statistics in MeshValmet
• Sample Err Mean:
• Sample Err STD:
• Face Err Mean:
• Face Err RMS:
• Hausdorff:
• MSD:
• MAD:
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
13
Alternative: CompareBYU
• Pros:
– Provides volume overlap of two surfaces
– Reads BYU-formatted input files - triangles or
quads
• Cons:
– Runs only on the command line and has no GUI
– Do not provide user-specified uniform sampling
• Only computes error on original vertices in the ‘BYU’
files (if the triangle tiles in the mesh are not all the same
size, the measures reported are not accurate enough)
2/11/11
MIDAG tutorial - MeshValmet, Christine Xu
14