Source-linked AI summary
Personalized Age Progression with Aging Dictionary
Xiangbo Shu, Jinhui Tang, Hanjiang Lai, Luoqi Liu, Shuicheng Yan
TL;DR
Age progression must preserve individual facial characteristics, but existing approaches either neglect those details or require difficult-to-collect long-term sequences. This paper learns age-group dictionaries from neighboring-age face pairs while separating aging and personalized layers. Experiments report advantages in personalized aging and gains in cross-age face verification.
Problem
Personalized age progression must retain individual facial characteristics, while prototype methods neglect such details and modeling approaches require dense long-term sequences that are difficult to collect.
Source: §1 · pp. 1–2Method
Results
Experiments report advantages over prior methods in personalized aging and gains in cross-age verification; 45.35% of ratings favored ours, versus 36.45% favoring prior works.
Source: §1 · p. 2, §4.3 · p. 7, §4.4 · p. 8Takeaways & Limitations
The authors report that synthesized aging faces can effectively mitigate age-gap effects in cross-age face verification.
Source: §4.4 · p. 8
Abstract
from arXiv · showhide
In this paper, we aim to automatically render aging faces in a personalized way. Basically, a set of age-group specific dictionaries are learned, where the dictionary bases corresponding to the same index yet from different dictionaries form a particular aging process pattern cross different age groups, and a linear combination of these patterns expresses a particular personalized aging process. Moreover, two factors are taken into consideration in the dictionary learning process. First, beyond the aging dictionaries, each subject may have extra personalized facial characteristics, e.g. mole, which are invariant in the aging process. Second, it is challenging or even impossible to collect faces of all age groups for a particular subject, yet much easier and more practical to get face pairs from neighboring age groups. Thus a personality-aware coupled reconstruction loss is utilized to learn the dictionaries based on face pairs from neighboring age groups. Extensive experiments well demonstrate the advantages of our proposed solution over other state-of-the-arts in term of personalized aging progression, as well as the performance gain for cross-age face verification by synthesizing aging faces.
1. Introduction
The paper targets personalized age progression that preserves invariant facial traits while avoiding reliance on difficult-to-collect long-term subject sequences. It learns age-group dictionaries from linked neighboring-age pairs using a personality-aware coupled reconstruction loss.
1. Introduction
- Prototype methods do not preserve personal details, while modeling methods do not specifically account for them and require dense long-term sequences that are difficult to collect.The paper motivates short-term sequences as a more feasible training source.Source: §1 · pp. 1–2
- Age-group dictionaries align same-index bases into aging patterns, while an individual face is split into aging and personalized layers that are recombined for synthesis.A sparse linear combination represents the aging layer; the personalized layer retains characteristics such as a mole.Source: §1 · p. 2
- Short-term face pairs from neighboring age groups jointly train the dictionaries because their links connect age groups and share sparse representations across each pair.The coupled reconstruction loss accounts for personalized details while learning the dictionaries.Source: §1 · p. 2
- Experiments report advantages over prior methods in personalized aging and performance gains for cross-age face verification using synthesized faces.The paper presents the approach as preserving personalized facial characteristics while requiring only available short-term face pairs.Source: §1 · p. 2
2. Related Work
Related work covers prototyping-based methods that transfer age differences between prototypes and modeling-based methods that synthesize facial shape and texture.
2. Related Work
- Prototyping-based methods transfer differences between age-group prototypes into an individual face; early work includes Burt et al. and later methods.Source: §2 · p. 2
- Prototyping-based methods synchronize aging speed and do not preserve personal characteristics, producing similar aging results across many faces.Source: §2 · p. 2
- Modeling-based methods synthesize shape and texture but generally need sufficient data for long-term aging; Suo et al. learn long-term patterns from short-term databases.Their approach uses a concatenational graph evolution aging model.Source: §2 · p. 2
3. Personalized Age Progression with Aging Dictionary
The method learns age-group dictionaries jointly from short-term neighboring-age face pairs, while separating age-related structure from invariant personalized details. At inference, it synthesizes one neighboring age at a time and feeds each result forward.
3.1. Overview of Our Framework
- The offline phase collects short-term face pairs, trains age-specific dictionaries jointly, and the online phase renders an input face into the nearest neighboring age group.Source: §3.1 · pp. 2–3, Figure 2 · p. 3
3.2. Formulation
- Each age group gets a dictionary whose bases model age-specific characteristics, while a personalized layer captures invariant details such as moles or birthmarks.Source: §3.2 · pp. 3–4
- A common sparse coefficient links same-person faces in neighboring age groups, allowing short-term pairs to train dictionaries jointly without long-term sequences.Source: §3.2 · p. 4
3.3. Optimization Procedure
- Dictionary learning alternates updates of coefficients, dictionaries, and personalized layers; the coefficient step is a Lasso problem and each dictionary update has a closed-form solution.Source: §3.3 · p. 4
3.4. Age Progression Synthesis
- For synthesis, the method optimizes a shared coefficient and personalized layer for the nearest age group, then feeds each rendered face into the next group.Source: §3.4 · p. 5
- The synthesis initializes from average faces and repeats rendering three times to obtain generally invariable, desired aging faces.Source: §3.4 · p. 5
4. Experiments
Experiments compare personalized aging against ground truth and prior methods, then test whether synthetic aging improves cross-age face verification. The reported results favor the proposed method in both evaluations.
4.1. Implementation Details
- Training used 1,600 neighboring-age face pairs, split evenly by gender; the objective converged after about 30 iterations with λ = 0.01 and γ = 0.1.The study also describes collecting dense short-term aging pairs from web photos and available databases.Source: §4.1 · pp. 5–6
4.2. Qualitative Comparison with Ground Truth
- Against ground truth, outputs better matched texture, facial outline, and aging speed than FT Demo and IAAP in the reported examples.The comparison spans younger inputs aged to adulthood or old age, and adult inputs aged to old age.Source: §4.2 · pp. 6–7, Figure 4 · p. 7
4.3. Quantitative Comparison with Prior Works
- 45.35% judged ours better, 36.45% prior works better, and 18.20% comparable across 12,300 ratings.The ratings came from 246 comparison groups.Source: §4.3 · p. 7, Figure 5 · p. 8
4.4. Evaluation on Cross-Age Face Verification
- Our Synthetic Pairs achieved lower ERR than Original Pairs and IAAP Synthetic Pairs; estimated-age variants outperformed label-based variants.The authors report that synthetic aging can mitigate the age-gap effect in cross-age verification.Source: §4.4 · pp. 7–8
5. Conclusions and Future Work
Age-group dictionaries encode aging patterns whose linear combinations express personalized aging, and neighboring-age pairs train the coupled dictionaries. The method renders a personalized sequence step by step; future work considers bilevel optimization for dictionary learning.
5. Conclusions and Future Work
- Age-group dictionaries form cross-age aging patterns, whose linear combinations express an individual's aging process.The dictionaries are trained jointly using younger- and older-age faces of the same people from neighboring age groups, with common sparse coefficients excluding the personalized layer.Source: §5 · p. 8
- Figure 6 presents face-verification pair settings and performance for Our Synthetic Pairs and IAAP Synthetic Pairs, with actual-age and estimated-age variants.Our Synthetic Pairs use the proposed method; IAAP Synthetic Pairs use IAAP [13].Source: Figure 6 · p. 8
- For an input face, the method renders a personalized aging sequence from the current age to future ages step by step.The learned aging dictionaries support this progression.Source: §5 · p. 8