1 Introduction
2 CNN-based age estimation
2.1 Single CNN-based age estimation
Layer | #Kernels | Size/stride | #Output nodes |
---|---|---|---|
conv1 | 81 | 5 × 5/1 | – |
pool1 | – | 3 × 3/2 | – |
conv1 | 45 | 7 × 7/1 | – |
pool2 | – | 2 × 2/2 | – |
fc3 | – | – | 1024 |
fc4_age | – | – | 1 |
2.2 Multistage CNN-based age estimation
2.2.1 Learning method
2.2.2 Definition of age-group classes
3 Performance evaluation
3.1 Database
Age group | Male | Female | Total |
---|---|---|---|
0–5 | 418 | 412 | 830 |
6–10 | 2349 | 2391 | 4740 |
11–15 | 2108 | 2265 | 4373 |
16–60 | 10,269 | 10,874 | 21,143 |
Over 60 | 452 | 385 | 837 |
Total | 15,596 | 16,327 | 31,923 |
Age group | Male | Female | Total |
---|---|---|---|
0–5 | 405 | 425 | 830 |
6–10 | 2351 | 2389 | 4740 |
11–15 | 2119 | 2254 | 4373 |
16–60 | 10,220 | 10,923 | 21,143 |
Over 60 | 402 | 435 | 837 |
Total | 15,497 | 16,426 | 31,923 |
3.2 Training settings
Ground truth | Total | |||
---|---|---|---|---|
Male | Female | |||
Predicted | Male | 15,470 | 165 | 15,635 |
Female | 126 | 16,162 | 16,288 | |
Total | 15,596 | 16,327 | 31,923 |
Ground truth | Total | ||||||
---|---|---|---|---|---|---|---|
0–5 | 6–10 | 11–15 | 16–60 | Over 60 | |||
(a) Predicted as male | |||||||
Predicted | 0–5 | 404 | 74 | 1 | 0 | 0 | 479 |
6–10 | 10 | 2200 | 161 | 2 | 0 | 2373 | |
11–15 | 0 | 76 | 1459 | 77 | 5 | 1617 | |
16–60 | 0 | 7 | 485 | 10,120 | 119 | 10,731 | |
Over 60 | 0 | 0 | 3 | 73 | 359 | 435 | |
Total | 414 | 2357 | 2109 | 10,272 | 483 | 15,635 | |
(b) Predicted as female | |||||||
Predicted | 0–5 | 405 | 19 | 0 | 0 | 0 | 424 |
6–10 | 11 | 2316 | 57 | 1 | 0 | 2385 | |
11–15 | 0 | 35 | 1919 | 83 | 0 | 2037 | |
16–60 | 0 | 13 | 288 | 10,776 | 40 | 11,117 | |
Over 60 | 0 | 0 | 0 | 11 | 314 | 325 | |
Total | 416 | 2,383 | 2,264 | 10,871 | 354 | 16,288 |
3.3 Evaluation method
3.4 Comparison with existing methods not based on CNNs
3.5 Sequential multi-CNN vs. parallel multi-CNN
Method | MAE [years old] | SD [years old] |
---|---|---|
Sequential multi-CNN (ours) |
5.84
|
6.50
|
Parallel multi-CNN
| 6.23 | 6.61 |
4 Discussion
4.1 Distribution of the estimated ages corresponding to the actual age
4.2 Order of learning tasks in Sequential multi-CNN
-
Age is trained last because age estimation is the target task.
-
Age group is trained second to the last because age group has a stronger relationship with age.
-
Gender is trained first because gender is easier to recognize than age group.
Ground truth | Total | |||
---|---|---|---|---|
Male | Female | |||
Predicted | Male | 15,581 | 413 | 15,594 |
Female | 316 | 16,013 | 16,329 | |
Total | 15,497 | 16,426 | 31,923 |
Ground truth | Total | ||||||
---|---|---|---|---|---|---|---|
0–5 | 6–10 | 11–15 | 16–60 | Over 60 | |||
(a) Male | |||||||
Predicted | 0–5 | 266 | 212 | 1 | 1 | 0 | 480 |
6–10 | 36 | 1912 | 350 | 4 | 9 | 2311 | |
11–15 | 0 | 213 | 1209 | 120 | 15 | 1557 | |
16–60 | 0 | 17 | 604 | 9969 | 244 | 10,834 | |
Over 60 | 0 | 0 | 9 | 137 | 266 | 412 | |
Total | 302 | 2354 | 2173 | 10,231 | 534 | 15,594 | |
(b) Female | |||||||
Predicted | 0–5 | 334 | 132 | 1 | 0 | 0 | 467 |
6–10 | 108 | 2178 | 317 | 17 | 0 | 2620 | |
11–15 | 0 | 182 | 1099 | 271 | 3 | 1555 | |
16–60 | 0 | 66 | 756 | 10,367 | 257 | 11,446 | |
Over 60 | 0 | 2 | 5 | 81 | 153 | 241 | |
Total | 442 | 2560 | 2178 | 10,736 | 413 | 16,329 |
4.3 Difference of accuracy between male and female
Gender | MAE [years old] | SD [years old] |
---|---|---|
Male |
5.60
|
6.19
|
Female | 6.07 | 6.77 |