Supplementary Material · in each panel is the original face image and the subsequent ... C.-S....
6
Supplementary Material Hongyu Yang 1 Di Huang 1* Yunhong Wang 1 Anil K. Jain 2 1 Beijing Advanced Innovation Center for Big Data and Brain Computing, Beihang University, China 2 Michigan State University, USA {hongyuyang, dhuang, yhwang}@buaa.edu.cn [email protected] Abstract Supplementary materials to the main paper. 1. Detailed Network Architecture The architectures of the networks G and D are shown in Table 1 and Table 2. 2. Database Properties We mainly use MORPH [3] and CACD [2] for training and validation. FG-NET [1] is also adopted for testing to make a fair comparison with prior work. The databases used in this paper are with different properties (see Table 3 and Fig. 1). 3. Additional Experimental Results 3.1. Additional Age Progression Results Additional synthesized faces achieved on CACD and MORPH are provided in Figs. 2 and 3. The first image in each panel is the original face image and the subsequent 3 images are the age progressed visualizations for that sub- ject in the [31- 40], [41-50] and 50+ age clusters. Although the examples cover a wide range of population in terms of race, gender, pose, makeup and expression, visually plausi- ble and convincing aging effects are achieved. Our method is not only shown to be effective but also robust to the other variations. 3.2. Rejuvenating Simulation Results The proposed method can also be applied for face reju- venating simulation. In this experiment, all the test faces come from the people older than 30 years old, and they are transformed to the age bracket of below 30 years old. Ex- ample rejuvenating visualizations are shown in Figs. 4 and 5. As can be seen, this operation tightens the face skin, and the hair becomes thick and luxuriant as expected. * Corresponding Author 3.3. Additional comparison to the one-pathway dis- criminator Figs. 6 and 7 provide more example faces compared with the one-pathway discriminator. Rejuvenating results are considered. References [1] The FG-NET Aging Database. http://www. fgnet.rsunit.com/ and http://www-prima. inrialpes.fr/FGnet/. 1, 2 [2] B.-C. Chen, C.-S. Chen, and W. H. Hsu. Face recognition and retrieval using cross-age reference coding with cross-age celebrity dataset. IEEE TMM, 17(6):804–815, 2015. 1, 2 [3] K. Ricanek and T. Tesafaye. Morph: A longitudinal image database of normal adult age-progression. In FG, pages 341– 345, Apr. 2006. 1, 2 1
Transcript of Supplementary Material · in each panel is the original face image and the subsequent ... C.-S....
Supplementary Material
Hongyu Yang1 Di Huang1lowast Yunhong Wang1 Anil K Jain2
1Beijing Advanced Innovation Center for Big Data and Brain Computing Beihang University China2Michigan State University USA
hongyuyang dhuang yhwangbuaaeducn jaincsemsuedu
Abstract
Supplementary materials to the main paper
1 Detailed Network ArchitectureThe architectures of the networks G and D are shown in
Table 1 and Table 2
2 Database PropertiesWe mainly use MORPH [3] and CACD [2] for training
and validation FG-NET [1] is also adopted for testing tomake a fair comparison with prior work The databases usedin this paper are with different properties (see Table 3 andFig 1)
3 Additional Experimental Results31 Additional Age Progression Results
Additional synthesized faces achieved on CACD andMORPH are provided in Figs 2 and 3 The first imagein each panel is the original face image and the subsequent3 images are the age progressed visualizations for that sub-ject in the [31- 40] [41-50] and 50+ age clusters Althoughthe examples cover a wide range of population in terms ofrace gender pose makeup and expression visually plausi-ble and convincing aging effects are achieved Our methodis not only shown to be effective but also robust to the othervariations
32 Rejuvenating Simulation Results
The proposed method can also be applied for face reju-venating simulation In this experiment all the test facescome from the people older than 30 years old and they aretransformed to the age bracket of below 30 years old Ex-ample rejuvenating visualizations are shown in Figs 4 and5 As can be seen this operation tightens the face skin andthe hair becomes thick and luxuriant as expected
lowast Corresponding Author
33 Additional comparison to the one-pathway dis-criminator
Figs 6 and 7 provide more example faces comparedwith the one-pathway discriminator Rejuvenating resultsare considered
References[1] The FG-NET Aging Database httpwww
fgnetrsunitcom and httpwww-primainrialpesfrFGnet 1 2
[2] B-C Chen C-S Chen and W H Hsu Face recognitionand retrieval using cross-age reference coding with cross-agecelebrity dataset IEEE TMM 17(6)804ndash815 2015 1 2
[3] K Ricanek and T Tesafaye Morph A longitudinal imagedatabase of normal adult age-progression In FG pages 341ndash345 Apr 2006 1 2
1
Table 1 Generator architecture
Layer conv convdarr convdarr res res res res deconvuarr deconvuarr deconv
Kernel 9 3 3 3 3 3 3 3 3 9Stride 1 2 2 1 1 1 1 2 2 1
Padding 4 1 1 2 2 2 2 1 1 4Outputs 32 64 128 128 128 128 128 64 32 3
Table 2 Discriminator architecture
Pathway Input Layers (denote as conv - ltouputgt kernel = 4 stride = 2 padding = 1 )
1 512 conv-512 conv-512 conv-12 256 conv-512 conv-512 conv-512 conv-13 128 conv-256 conv-512 conv-512 conv-512 conv-14 64 conv-128 conv-256 conv-512 conv-512 conv-512 conv-1
Table 3 Statistics of face aging databases used for evaluation
Database Number ofimages
Number ofsubjects
Number ofimages per subject
Time lapseper subject (years)
Age span(years old)
Average age(years old)
MORPH [3] 52099 12938 1 - 53 (avg 403) 0 - 33 (avg 162) 16 - 77 3307CACD [2] 163446 2000 22 - 139 (avg 8172) 7 - 9 (avg 899) 14 - 62 3803
FG-NET [1] 1002 82 6 - 18 (avg 1222) 11 - 54 (avg 2780) 0 - 69 1584
(a) MORPH
1 -10
11 -20
21 - 30
31 - 40
41 - 50
51 - 60
61 + 050
570
2129
2921
2755
1575
000
(b) CACD
1 -10
11 -20
21 - 30
31 - 40
41 - 50
51 - 60
61 + 096
1743
2436
2642
2493
590
000
(c) FGNET
1 -10
11 -20
21 - 30
31 - 40
41 - 50
51 - 60
61 + 070
140
389
689
1427
3184
4102
Figure 1 Age distributions of (a) MORPH (b) CACD and (c) FGNET
30 years old 26 years old 30 years old
23 years old 22 years old 30 years old
28 years old 25 years old 28 years old
17 years old 29 years old 28 years old
26 years old 30 years old 23 years old
29 years old 27 years old 26 years old
27 years old 29 years old 27 years old
17 years old 28 years old 16 years old
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
Figure 2 Additional aging effects obtained on the CACD databases for 24 different subjects The first image in each panel is the originalface image and the subsequent 3 images are the age progressed visualizations for that subject in the [31- 40] [41-50] and 50+ age clusters
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
25 years old 26 years old 28 years old
26 years old 23 years old 27 years old
29 years old 29 years old 24 years old
28 years old 30 years old 29 years old
17 years old 25 years old 28 years old
30 years old 30 years old 20 years old
28 years old 26 years old 28 years old
26 years old 28 years old 29 years old
Figure 3 Additional aging effects obtained on the MORPH databases for 24 different subjects
55 years old 59 years old 57 years old 57 years old 57 years old 52 years old
48 years old 57 years old 49 years old 56 years old 56 years old 55 years old
55 years old 48 years old 51 years old 47 years old 56 years old 56 years old
47 years old 55 years old 52 years old 52 years old 55 years old 56 years old
Figure 4 Rejuvenating results achieved on the CACD database for 24 different subjects The first image in each panel is the original faceimage and the second is the corresponding rejuvenating result
49 years old 50 years old 46 years old 54 years old 43 years old 48 years old
44 years old 48 years old 45 years old43 years old 42 years old 44 years old
43 years old 50 years old 44 years old 43 years old 46 years old 47 years old
44 years old 41 years old 43 years old 42 years old56 years old 42 years old
Figure 5 Rejuvenating results achieved on the MORPH database for 24 different subjects
Test face
One-Pathway Discriminator
Proposed
18 26 20 30 27 55 58 51 51 57
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 6 Additional visual comparison to the one-pathway discriminator on the MORPH database
Test face
One-Pathway Discriminator
Proposed
30 29 30 30 26 53 49 52 52 49
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 7 Additional visual comparison to the one-pathway discriminator on the CACD database
Table 1 Generator architecture
Layer conv convdarr convdarr res res res res deconvuarr deconvuarr deconv
Kernel 9 3 3 3 3 3 3 3 3 9Stride 1 2 2 1 1 1 1 2 2 1
Padding 4 1 1 2 2 2 2 1 1 4Outputs 32 64 128 128 128 128 128 64 32 3
Table 2 Discriminator architecture
Pathway Input Layers (denote as conv - ltouputgt kernel = 4 stride = 2 padding = 1 )
1 512 conv-512 conv-512 conv-12 256 conv-512 conv-512 conv-512 conv-13 128 conv-256 conv-512 conv-512 conv-512 conv-14 64 conv-128 conv-256 conv-512 conv-512 conv-512 conv-1
Table 3 Statistics of face aging databases used for evaluation
Database Number ofimages
Number ofsubjects
Number ofimages per subject
Time lapseper subject (years)
Age span(years old)
Average age(years old)
MORPH [3] 52099 12938 1 - 53 (avg 403) 0 - 33 (avg 162) 16 - 77 3307CACD [2] 163446 2000 22 - 139 (avg 8172) 7 - 9 (avg 899) 14 - 62 3803
FG-NET [1] 1002 82 6 - 18 (avg 1222) 11 - 54 (avg 2780) 0 - 69 1584
(a) MORPH
1 -10
11 -20
21 - 30
31 - 40
41 - 50
51 - 60
61 + 050
570
2129
2921
2755
1575
000
(b) CACD
1 -10
11 -20
21 - 30
31 - 40
41 - 50
51 - 60
61 + 096
1743
2436
2642
2493
590
000
(c) FGNET
1 -10
11 -20
21 - 30
31 - 40
41 - 50
51 - 60
61 + 070
140
389
689
1427
3184
4102
Figure 1 Age distributions of (a) MORPH (b) CACD and (c) FGNET
30 years old 26 years old 30 years old
23 years old 22 years old 30 years old
28 years old 25 years old 28 years old
17 years old 29 years old 28 years old
26 years old 30 years old 23 years old
29 years old 27 years old 26 years old
27 years old 29 years old 27 years old
17 years old 28 years old 16 years old
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
Figure 2 Additional aging effects obtained on the CACD databases for 24 different subjects The first image in each panel is the originalface image and the subsequent 3 images are the age progressed visualizations for that subject in the [31- 40] [41-50] and 50+ age clusters
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
25 years old 26 years old 28 years old
26 years old 23 years old 27 years old
29 years old 29 years old 24 years old
28 years old 30 years old 29 years old
17 years old 25 years old 28 years old
30 years old 30 years old 20 years old
28 years old 26 years old 28 years old
26 years old 28 years old 29 years old
Figure 3 Additional aging effects obtained on the MORPH databases for 24 different subjects
55 years old 59 years old 57 years old 57 years old 57 years old 52 years old
48 years old 57 years old 49 years old 56 years old 56 years old 55 years old
55 years old 48 years old 51 years old 47 years old 56 years old 56 years old
47 years old 55 years old 52 years old 52 years old 55 years old 56 years old
Figure 4 Rejuvenating results achieved on the CACD database for 24 different subjects The first image in each panel is the original faceimage and the second is the corresponding rejuvenating result
49 years old 50 years old 46 years old 54 years old 43 years old 48 years old
44 years old 48 years old 45 years old43 years old 42 years old 44 years old
43 years old 50 years old 44 years old 43 years old 46 years old 47 years old
44 years old 41 years old 43 years old 42 years old56 years old 42 years old
Figure 5 Rejuvenating results achieved on the MORPH database for 24 different subjects
Test face
One-Pathway Discriminator
Proposed
18 26 20 30 27 55 58 51 51 57
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 6 Additional visual comparison to the one-pathway discriminator on the MORPH database
Test face
One-Pathway Discriminator
Proposed
30 29 30 30 26 53 49 52 52 49
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 7 Additional visual comparison to the one-pathway discriminator on the CACD database
30 years old 26 years old 30 years old
23 years old 22 years old 30 years old
28 years old 25 years old 28 years old
17 years old 29 years old 28 years old
26 years old 30 years old 23 years old
29 years old 27 years old 26 years old
27 years old 29 years old 27 years old
17 years old 28 years old 16 years old
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
Figure 2 Additional aging effects obtained on the CACD databases for 24 different subjects The first image in each panel is the originalface image and the subsequent 3 images are the age progressed visualizations for that subject in the [31- 40] [41-50] and 50+ age clusters
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
25 years old 26 years old 28 years old
26 years old 23 years old 27 years old
29 years old 29 years old 24 years old
28 years old 30 years old 29 years old
17 years old 25 years old 28 years old
30 years old 30 years old 20 years old
28 years old 26 years old 28 years old
26 years old 28 years old 29 years old
Figure 3 Additional aging effects obtained on the MORPH databases for 24 different subjects
55 years old 59 years old 57 years old 57 years old 57 years old 52 years old
48 years old 57 years old 49 years old 56 years old 56 years old 55 years old
55 years old 48 years old 51 years old 47 years old 56 years old 56 years old
47 years old 55 years old 52 years old 52 years old 55 years old 56 years old
Figure 4 Rejuvenating results achieved on the CACD database for 24 different subjects The first image in each panel is the original faceimage and the second is the corresponding rejuvenating result
49 years old 50 years old 46 years old 54 years old 43 years old 48 years old
44 years old 48 years old 45 years old43 years old 42 years old 44 years old
43 years old 50 years old 44 years old 43 years old 46 years old 47 years old
44 years old 41 years old 43 years old 42 years old56 years old 42 years old
Figure 5 Rejuvenating results achieved on the MORPH database for 24 different subjects
Test face
One-Pathway Discriminator
Proposed
18 26 20 30 27 55 58 51 51 57
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 6 Additional visual comparison to the one-pathway discriminator on the MORPH database
Test face
One-Pathway Discriminator
Proposed
30 29 30 30 26 53 49 52 52 49
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 7 Additional visual comparison to the one-pathway discriminator on the CACD database
Test face 31 - 40 41 - 50 50+ Test face 31 - 40 41 - 50 50+Test face 31 - 40 41 - 50 50+
25 years old 26 years old 28 years old
26 years old 23 years old 27 years old
29 years old 29 years old 24 years old
28 years old 30 years old 29 years old
17 years old 25 years old 28 years old
30 years old 30 years old 20 years old
28 years old 26 years old 28 years old
26 years old 28 years old 29 years old
Figure 3 Additional aging effects obtained on the MORPH databases for 24 different subjects
55 years old 59 years old 57 years old 57 years old 57 years old 52 years old
48 years old 57 years old 49 years old 56 years old 56 years old 55 years old
55 years old 48 years old 51 years old 47 years old 56 years old 56 years old
47 years old 55 years old 52 years old 52 years old 55 years old 56 years old
Figure 4 Rejuvenating results achieved on the CACD database for 24 different subjects The first image in each panel is the original faceimage and the second is the corresponding rejuvenating result
49 years old 50 years old 46 years old 54 years old 43 years old 48 years old
44 years old 48 years old 45 years old43 years old 42 years old 44 years old
43 years old 50 years old 44 years old 43 years old 46 years old 47 years old
44 years old 41 years old 43 years old 42 years old56 years old 42 years old
Figure 5 Rejuvenating results achieved on the MORPH database for 24 different subjects
Test face
One-Pathway Discriminator
Proposed
18 26 20 30 27 55 58 51 51 57
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 6 Additional visual comparison to the one-pathway discriminator on the MORPH database
Test face
One-Pathway Discriminator
Proposed
30 29 30 30 26 53 49 52 52 49
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 7 Additional visual comparison to the one-pathway discriminator on the CACD database
55 years old 59 years old 57 years old 57 years old 57 years old 52 years old
48 years old 57 years old 49 years old 56 years old 56 years old 55 years old
55 years old 48 years old 51 years old 47 years old 56 years old 56 years old
47 years old 55 years old 52 years old 52 years old 55 years old 56 years old
Figure 4 Rejuvenating results achieved on the CACD database for 24 different subjects The first image in each panel is the original faceimage and the second is the corresponding rejuvenating result
49 years old 50 years old 46 years old 54 years old 43 years old 48 years old
44 years old 48 years old 45 years old43 years old 42 years old 44 years old
43 years old 50 years old 44 years old 43 years old 46 years old 47 years old
44 years old 41 years old 43 years old 42 years old56 years old 42 years old
Figure 5 Rejuvenating results achieved on the MORPH database for 24 different subjects
Test face
One-Pathway Discriminator
Proposed
18 26 20 30 27 55 58 51 51 57
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 6 Additional visual comparison to the one-pathway discriminator on the MORPH database
Test face
One-Pathway Discriminator
Proposed
30 29 30 30 26 53 49 52 52 49
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 7 Additional visual comparison to the one-pathway discriminator on the CACD database
Test face
One-Pathway Discriminator
Proposed
18 26 20 30 27 55 58 51 51 57
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 6 Additional visual comparison to the one-pathway discriminator on the MORPH database
Test face
One-Pathway Discriminator
Proposed
30 29 30 30 26 53 49 52 52 49
(a) Aging Simulation (b) Rejuvenating Simulation
Figure 7 Additional visual comparison to the one-pathway discriminator on the CACD database
