5.4 Results and Analysis
5.4.2 Comparative Text and Image Features
Among the three criteria, we utilized the high correlation achieving WG and WUG for cross-modal visual attribute comparison. The approach described in subsection 5.3.2 is utilized to obtain the comparative text and image features with respect to the free-viewing attention. Though, in total, the number of comparisons achieved is equal to the minimum feature count of text and image, the majority of them results in very low canonical loading.
Consequently, we limit to compare the Top-3 visual features based on associated canonical loadings.
The font-family (‘open sans’) of the text iscomparable with the color histogram (mid-level blue color component) of the image along the most prominent canonical direction for WG grouping as shown in Figure 5.5. The text feature achieved a canonical loading of 1.00 and the image feature achieved a canonical loading of 0.46. To note, in the histogram feature denoted as ‘hist_B_6_182’, ‘B’ (from R, G, B, gray) represents the blue color,
‘6’ (from 1 to 8) represents the histogram bin number, and the number (182) denotes a
5.4. RESULTS AND ANALYSIS representative color component value (∈[0,255]) for the respective bin. Further, the font- family (‘arial’) of text is comparable to the color histogram of the image (mid-level blue color histogram). In aggregation, the FAMLIY of the text is comparable to the HISTOGRAM of images in describing the free-viewing user attention on interfaces. This comparison is further strengthened by the very high squared ρ2max = 0.9896 along the first canonical direction. With respect to the Top-2 canonical loadings (withρ2max>0.8), the FAMILY, COLOR, SPACE features of text are as influential as HISTOGRAM, CONTRAST HIST, and POSITION features of the images in establishing the correlation between both the modalities.
diff_R_var 0.30 CONTRAST COMPRE
0.9896
Decreasing squared canonical correlations
COLOR
FAMILY 0.35 0.25
FAMILY 1.00 R_background.color
arial
HISTOGRAM hist_B_6_182 0.46
HISTOGRAM hist_B_5_146 0.38
opensans
hist_R_6_182 0.40 HISTOGRAM
0.8755
FAMILY SPACE
0.53
0.68 FAMILY
0.49
arial line.height
CONTRAST HIST 0.68 diff_hist_R_5_146
HISTOGRAM 0.41 hist_R_5_146
helvetica
hist_G_5_146 0.20 HISTOGRAM
0.8259
COLOR COLOR
0.77
0.82 COLOR
0.50
opacity_color opacity_border.left
POSITION 0.29 rect.left
HISTOGRAM 0.20 hist_B_6_182 opacity_border.bottom
0.23 hist_R_7_219
HISTOGRAM
0.8217
FAMILY COLOR
0.49
0.67 COLOR
0.45
helvetica opacity_color
HISTOGRAM hist_G_7_219 0.27
HISTOGRAM hist_gray_7_219 0.24 opac_border.left.color
decreasing im age canonical loadings decre
asin g tex
t can onica
l loadin gs
Figure 5.5: Interpretative visualization of WGcanonical loadings. The decreasing intensity and size of the ball (moving from left to right) indicates the decreasing prominence (inscribed with ρ2max) of the corresponding canonical direction. At each position, text feature cuboids are placed above and image feature cuboids are placed below the ball in the decreasing order of canonical loadings. Each cuboid consists of feature name (on front), feature group (on top), and corresponding canonical loading (on side).
As shown in Figure 5.6 for WUG, along the most prominent canonical direction, the FAMILY feature achieved the highest canonical loading of 1.00 while the CONTRAST COMPRE feature achieved 0.21. That is, a significant difference exists between both the canonical loading scores. On the other hand, along the second most canonical direction (withρ2max= 0.9758), the FAMILY group achieved comparable canonical loadings with that of the image’s HISTOGRAM features. Similarly, FAMILY and HISTOGRAM features are comparable in achieving the correlation maximization along the third canonical direction. In contrast, FAMILY features are comparable to the CONTRAST HIST features in achieving
0.8003
FAMILY – 0.38 (proxima_nova_rgregular)
FAMILY – 0.51 (verdana)
SIZE – 0.31 (font.size)
FAMILY – 0.39 (proxima_nova_rgregular)
FAMILY – 0.17 (verdana)
FAMILY – 0.25 (arial)
FAMILY – 0.13 (helveticaneue)
NT CO ST RA OM C E – PR 17 0.
ff_ (di vg) G_a
SIZE – 0.12 (font.size)
0.9890
TO HIS AM GR 0.1 – 9
ff_ (di t_R his 255 _8_
)
SPACE – 0.40 (margin.bottom) FAMILY – 1.00 (helveticaneuelight)
NT CO ST RA OM C E – PR 21 0.
ff_ (di ura Sat n_m tio ) ean
TO HIS AM GR 0.3 – 5
st_ (hi _10 G_4 9)
0.9758
TO HIS AM GR 0.3 –
6 2) _18 B_6 st_ (hi
FAMILY – 0.18 (opensans) FAMILY – 0.98
(lucidagrande) TO HIS AM GR 0.4 –
0 6) _14 B_5 st_ (hi
TO HIS AM GR 0.5 – 8
st_ (hi _0) B_1
TO HIS AM GR 0.5 –
9 ) _36 R_2 st_ (hi
FAMILY – 0.43 (ptsans) FAMILY – 0.86 (helveticaneue) TO HIS AM GR 0.6 –
1 2) _18 G_6 st_ (hi
0.9550
NT CO ST RA IS H 0. T – 31
ff_ (di t_B his 182 _6_
)
NT CO ST RA IS H 0. T – 40
ff_ (di t_G his 182 _6_
)
FAMILY – 0.49 (arial) FAMILY – 0.94
(opensans) NT CO ST RA IS H 0. T – 44
ff_ (di t_g his _5_ ray ) 146
0.9192
NT CO ST RA IS H 0. T – 31
ff_ (di t_R his 109 _4_
)
NT CO ST RA IS H 0. T – 44
ff_ (di t_G his 146 _5_
)
FAMILY – 0.41 (verdana) FAMILY – 0.52
(helvetica) NT CO ST RA IS H 0. T – 51
ff_ (di t_R his 146 _5_
)
0.8494
MP CO HE RE IV NS 0.2 E–
9
atu (S
ion rat
ean _m )
NT CO ST RA OM C E – PR 31 0.
ff_ (di _m Hue ) ean
FAMILY – 0.64 (verdana) FAMILY – 0.72
(georgia) MP CO HE RE IV NS 0. E –
46 n) mea ue_ (H
0.8380
NT CO ST RA OM C E – PR 24 0.
ff_ (di vg) R_a
NT CO ST RA OM C E – PR 26 0.
ff_ (di
ina Lum
_av nce g)
FAMILY – 0.56 (georgia) FAMILY – 0.56
(arial) NT CO ST RA OM C E – PR 27 0.
ff_ (di vg) G_a
0.8031
TO HIS AM GR 0.1 – 8
st_ (hi _10 G_4 9)
TO HIS AM GR 0.1 –
9 6) _14 R_5 st_ (hi
FAMILY – 0.39 (serif) FAMILY – 0.74
(helvetica) NT CO ST RA IS H 0. T – 24
ff_ (di t_B his 36) _2_
Decreasing squared canonical correlations
ca no nic al d ire ctio n-1
ca no nic al d ire ctio n-2
ca no nic al d ire ctio n-3
ca no nic al d ire ctio n-4
ca no nic al d ire ctio n-5
ca no nic al
dire ctio n-6
ca no nic al
dir ect io n-7
ca no nic al
dire ctio n-8
decreasing text canonical loadings decreasing image canonical loadings Figure5.6:InterpretativevisualizationofWUGcanonicalloadings.Thedecreasingtreeheight(fromlefttoright)indicatesthedecreasing prominence(indicatedwithρ2 maxinthewaterballoonatthebottom)ofthecorrespondingcanonicaldirection.Foreachtree,textfeature leavesareplacedontheleftsideandimagefeatureleavesareplacedontherightsideofthetreetrunkfromtoptobottominthe decreasingorderofcanonicalloadings.Eachleafconsistsoffeaturename,featuregroup,andcorrespondingcanonicalloading.
5.4. RESULTS AND ANALYSIS
the fourth and fifth canonical direction. The latter feature comparison is of little interest as the corresponding ρ2max values as well as the canonical loadings are reducing. Overall, FAMILY features of text dominated the canonical loadings across the canonical directions with ρ2max >0.8 and are comparable to the attention associated with the HISTOGRAM (followed by CONTRAST HIST) features, as visualized in Figure 5.6.
In summary, from the interface idiosyncrasies constrained WGandWUG groupings,FAM- ILY features of text are comparable to and as influential as HISTOGRAM and CONTRAST HIST features of images in explaining the free-viewing user atten- tion (answer to research question R2). The low canonical correlation obtained for UG grouping indicates that the proposed method cannot enable the comparison of text and image features for the marginalization of user idiosyncrasies.