How Well Do Line Drawings Depict Shape?
Forrester ColeKevin Sanik
Doug DeCarloAdam FinkelsteinThomas FunkhouserSzymon Rusinkiewicz
Manish Singh
RutgersPrinceton
Line drawings
[Matisse 1932]
[Flaxman 1805]
[US Patent 378,973]
Line drawings
Occluding Contours Sharp creases
Line drawings
Ridges and Valleys Suggestive Contours[DeCarlo et al 2003]
Apparent Ridges[Judd et al 2007]
Occluding Contours Sharp creases
Assessing Line Drawings
• Goals– Artistry, abstraction, etc.– Leads to accurate perception of shape
Assessing Line Drawings
• Goals– Artistry, abstraction, etc.– Leads to accurate perception of shape
• Methodology– Qualitative (examples, comparison to artists)– Quantitative comparison to artists’ drawings– Direct measurement of perceived shape
Comparing models
Ridges and ValleysSuggestive Contours Apparent Ridges
Comparing models
Ridges and ValleysSuggestive Contours Apparent Ridges
Comparing models
Ridges and ValleysSuggestive Contours Apparent Ridges
Comparing models to artists
© Estate of Roy Lichtenstein Suggestive contours and
suggestive highlights [DeCarlo and Rusinkiewicz 2007]
“Golf Ball” [Lichtenstein 1962]
Comparing models to artists
• Argument for ridge-like features [Judd et al. 2007]
[Brancusi 1910][Matisse 1932]
Comparing models to artists
• Comparisons to drawings made under controlled conditions [Cole et al. 2008]
artist drawingapparent ridges suggestive contours
…
Comparing models to artists
• Comparisons to drawings made under controlled conditions [Cole et al. 2008]
artist drawingapparent ridges suggestive contours
…
dline( , ) rendering artist drawing
rendering
Comparing shapes
3D
d3D( , )perceived shape original shape
Measuring perceived shape
Local measurements of shape geometry• Gauge figure adjustment
[Koenderink et al. 1992]
Measuring perceived shape
Local measurements of shape geometry• Gauge figure adjustment
[Koenderink et al. 1992]• Studied shaded surfaces
and one artist line drawing[Koenderink et al. 1996]
Questions
Do artist and CG drawings effectively convey shape?– how accurate are they?– how do they compare to a shaded rendering?
Do different viewers perceive the same shape?When are particular line types most effective?
Study Methodology
1. Measure percepts– Both artist and CG drawings– Range of models– Many participants
2. Compare against ground truth– 3D shape and shaded image– Accuracy and precision
Orienting a Gauge
Example Session
Study Setup
All 12 models from [Cole et al. 2008]
Shaded
R. and V.
Sug. C.
App. R.
Artist’s
Study Setup
6 styles x 12 models - 2 duplicates = 70 prompts
Contours
Shaded
R. and V.
Sug. C.
App. R.
Artist’s
Study Setup
6 styles x 12 models - 2 duplicates = 70 prompts
Contours
Study Setup
70x 90 gauges / prompt
x 2 settings / opinion
prompts
≈ 100,000 settings
x 8 opinions / gauge
Study Setup
x 4 seconds / setting
111 hours
70x 90 gauges / prompt
x 2 settings / opinion
prompts
≈ 100,000 settings
x 8 opinions / gauge
So Much Data…
• Amazon Mechanical Turk to the rescue!• Turker sets 60 gauges, gets paid $0.20• Efficient even after throwing away garbage– “Garbage” is inconsistent data– About 80% of data is consistent
Data Collection
• 275,000 gauge settings• 4 models x 180 gauges + 8 models x 90 gauges• Each gauge 9 to 29 opinions, average 15• 560 different people
1 53 1051572092613133654174695215730
102030405060
AssignmentsCompleted
# Participants
Global Accuracy
Error from ground (accuracy)
Global Accuracy
Error from ground (accuracy) Distribution of errors for shaded
Finding:
On average, turkers did a good job
Aggregating Per-Gauge Data
What is the most representative direction?– “Mean” is most obvious choice– “Median” more robust to outliers
mean median
Global Accuracy and Precision
Error from Ground (Accuracy) Error from Median (Precision)
Results:
• Precision greater than accuracy• Accuracy varies with style, precision does not
Finding:
Peoples’ interpretations of shape are similar, even when those interpretations do not match
ground truth.
Question:
Where are the errors?
Accuracy by Model
Vertebra
Cervical
Tooth
Bumps
Lumpcloth
Pulley
Femur
Screwdriver
Twoboxcloth
Flange
Rockerarm
Cubehole
0 10 20 30 40
Shaded Best Drawing
Avg. Error (degrees)
Accuracy by Model
Vertebra
Cervical
Tooth
Bumps
Lumpcloth
Pulley
Femur
Screwdriver
Twoboxcloth
Flange
Rockerarm
Cubehole
0 10 20 30 40
Shaded Best Drawing
Avg. Error (degrees)
Gauge Visualization: Screwdriver
Contours Only
Artist’s Drawing
0 90Error (deg.)180 gauges
Local Errors: Screwdriver
Contours OnlyArtist’s Drawing
15 gauges, 5 pixel spacing 0 90Error (deg.)
Curvature: Screwdriver
Contours Only
Artist’s Drawing
Contours Only
Artist’s Drawing
Ground Truth
Zero Curvature
Gauge Visualization: Flange
Suggestive Contours
180 gauges
Ridges and Valleys
0 90Error (deg.)
Local Errors: Flange
Suggestive ContoursRidges and Valleys
15 gauges, 5 pixel spacing 0 90Error (deg.)
Curvature: Flange
Suggestive Contours
Ridges and Valleys
Ground Truth
R. and V.
Sug. Contours
Gauge Visualization: Rockerarm
Apparent Ridges
90 gauges
Ridges and Valleys
0 90Error (deg.)
Non-Local Effects: Rockerarm
-90 90Error Difference (deg)
Worse than RVBetter than RVApparent Ridges
Conclusions
• Different people interpret drawings similarly• Some drawings almost match shaded images• Line drawings vary in effectiveness– Errors can be traced to specific lines
Future Work
• More analysis of collected data– Towards interpretation model for lines
• Further investigation of study methodology
Data available at:http://lineshape.cs.princeton.edu
Thank You• Thanks to Andrew Van Sant and John Wilder• Support by NSF grants CCF-0347427, CCF-0541185,
CCF- 0702672, CCF-0702580, IIS-0511965, and IIS-0612231, and Google
• Models from Aim@Shape, VAKHUN, and Cyberware
Data available at:http://lineshape.cs.princeton.edu
Global Accuracy and Precision
Before bas-relief fitting
Error from Ground (Accuracy) Error from Median (Precision)
Global Accuracy and Precision
After bas-relief fitting
Error from Ground (Accuracy) Error from Median (Precision)
Bas-Relief Ambiguity
Ambiguity in perception of shaded shapes [Koenderink 2001]
= ?
Line Drawing Ambiguity
Line drawings are even less constrained
= ?