Experimental Design

1
Experimental Design

• Viewing Conditions 1. D50 room and 9300K CRT
  at 70cd/m2 2. D50 booth _at_ 600cd/m2 9300K
  CRT _at_ 70cd/m2 3. Source A room or booth and
  9300K CRT _at_ 80cd/m2 Room lights off for
  CRT viewing

2
Three of Six Scenes
Ballpark
Boy
Garden
3
Three of Six Scenes
Litehouse
Pier
Restaurant
4
Experimental Design
Models Studied Based on Pilot

• Revisions of CIECAM97s (avg. surr.)
• CAM97s2 (Li, Luo, Hunt)
• CAM97s3 (Li, Luo, Rigg, Hunt)
• CAMMDF (Fairchild)
• Hunt 96 (avg. surr.)
• RLAB (avg. surr.)
• LLAB (avg. surr.)
• CIECAT94LAB (Nayatani)
• CIE chromatic adaptation transform with a
  CIELAB-like color space

5
Data Analysis

• Multi-Dimensional Scaling
• - Outlier Identification
• - Dimensionality
• Duncan Test of Statistical Significance
• - Eliminate based on Circular Triads
• Comrey Constant Sum Rating

6
Multi-Dimensional Scaling Results
of Total Variance Explained by Each Dimension
Scene
Dim. 1
Dim. 12
Dim. 1,23
Dim. 1,2,34
Ballpark
72
81
89
94
Boy
74
86
94
98
56
75
89
94
Garden
85
96
Litehouse
59
91
Pier
54
81
91
96
91
96
98
Restaurant
83
7
Duncan Scaling Results
8
Duncan Scaling Results
9
Duncan Scaling Results
10
Duncan Scaling Results
11
Duncan Scaling Results
12
Duncan Scaling Results
13
Comrey Constant Sum Rating

• Rates both members of the softcopy pair as
  fractions of 100 based on how close an appearance
  match they are to the hardcopy print
• Results in a rating for each model and each
  observer
• Allows for parametric analysis ANOVA and Duncan
  test
• Yields 95 confidence limits in addition to
  statistical signicance

14
Duncan Scaling Results from CCS Ratings
15
Duncan Scaling Results from CCS Ratings
16
Duncan Scaling Results from CCS Ratings
17
Duncan Scaling Results from CCS Ratings
18
Duncan Scaling Results from CCS Ratings
19
Duncan Scaling Results from CCS Ratings
20
Conclusions for 2nd Viewing Condition

• Results are scene-dependent
• RLAB(3) LLAB(1) performed the best for 4/6
  scenes
• Revisions of CIECAM97s, especially CAMMDF,
  performed best for 2/6 scenes (Restaurant Boy)
• Revisions of CIECAM97s performed similarly
• Scenes where CIECAM97s revisions were not best
  featured poor softcopy reproduction of colors in
  the R-Y quadrant
• CIECAT94LAB performed worst for all scenes

21
Conclusions for 2nd Viewing Condition

• Hunt 96 performed second worst for all scenes
  except Restaurant Boy
• Duncan from CCS Ratings yielded similar model
  ranking results to Duncan results from paired
  comparison with the exception of the order for
  the CIECAM97s revisions
• Duncan from CCS Ratings tended to be less
  sensitive to statistical differences
• Duncan from CCS Ratings allowed us to separate
  the best models from the average and poor
  performers

22
Conclusions for 2nd Viewing Condition

• On average, the percentage of total variance
  explained by up to four dimensions is
• 66 for 1 dim. 83 for 2 dim. 92
  for 3 dim. 96 for 4 dim.

23
Conclusions from Both Experiments

• When hardcopy and softcopy viewing conditions
  differ only in chromaticity, a chromatic
  adaptation transform, preferably BFD, is enough.
  A color appearance model, preferably CIECAM97
  revision, is probably overkill
• When hardcopy and softcopy viewing conditions
  differ in chromaticity and luminance level, RLAB
  LLAB performed better for most scenes than any
  of the CIECAM97s revisions
• MDS results showed dimensionality to be similar
  for both viewing conditions

24
Conclusions from Both Experiments

• CCS Rating technique provided data for all models
  for each observer, but observers felt it was
  harder to rate models as fractions of 100 than it
  was to simply pick which softcopy model provided
  an image that was a closer match to the
  illuminated print
• CCS Rating is a preferred technique because the
  95 confidence limits allow the quantification of
  how much better one model performs over another

25
Future Work

• TC 1-27 wishes to complete its work
• The questions that remain regarding why RLAB
  LLAB performed better than CIECAM97s revisions
  when luminance level varied across media might
  better be addressed by CIE TC8-01