[Paper Review] Cross-Age LFW: A Database for Studying Cross-Age Face Recognition in Unconstrained Environments

Tianyue Zheng, Weihong Deng+1 more|arXiv (Cornell University)|Aug 28, 2017

Face recognition and analysis20 references304 citations

TL;DR

CALFW introduces a cross-age variant of LFW by selecting 3,000 positive pairs with large age gaps and negative pairs with the same gender and race, revealing a 10-17% drop in accuracy compared to LFW across baseline methods.

ABSTRACT

Labeled Faces in the Wild (LFW) database has been widely utilized as the benchmark of unconstrained face verification and due to big data driven machine learning methods, the performance on the database approaches nearly 100%. However, we argue that this accuracy may be too optimistic because of some limiting factors. Besides different poses, illuminations, occlusions and expressions, cross-age face is another challenge in face recognition. Different ages of the same person result in large intra-class variations and aging process is unavoidable in real world face verification. However, LFW does not pay much attention on it. Thereby we construct a Cross-Age LFW (CALFW) which deliberately searches and selects 3,000 positive face pairs with age gaps to add aging process intra-class variance. Negative pairs with same gender and race are also selected to reduce the influence of attribute difference between positive/negative pairs and achieve face verification instead of attributes classification. We evaluate several metric learning and deep learning methods on the new database. Compared to the accuracy on LFW, the accuracy drops about 10%-17% on CALFW.

Motivation & Objective

• Motivate a more realistic face verification benchmark by introducing aging-related intra-class variation.
• Create CALFW by augmenting LFW identities with large age gaps while controlling for gender and race in negatives.
• Maintain LFW's verification protocol to enable direct comparison and easy adoption by researchers.

Proposed method

• Construct CALFW by gathering images for LFW identities with the largest feasible age gaps.
• Detect, crop, align, and deduplicate faces; manually verify labels and landmarks for accuracy.
• Estimate age of each image using the Dex model and form training/testing pairs emphasizing maximal age gaps for positives.
• Form negative pairs from individuals of the same gender and race to minimize attribute differences.
• Evaluate baseline metric learning and deep learning methods on CALFW and compare with LFW to gauge difficulty.

Experimental results

Research questions

• RQ1 Does introducing large cross-age intra-class variation reduce face verification performance relative to LFW?
• RQ2 How do attribute-controlled negatives (same gender/race) affect verification difficulty on CALFW?
• RQ3 Do standard metric learning and CNN-based approaches retain performance when aging is explicitly modeled as intra-class variation?

Key findings

• CALFW increases intra-class aging variance by enlarging the positive-pair age gap to an average of 16.61 years (SD 10.78) versus 4.94 years (SD 4.24) in LFW.
• CALFW negatives are selected to have the same gender and race as positives to reduce attribute differences.
• Across baselines, CALFW reduces verification accuracy by about 10–17% compared to LFW (e.g., ITML: 82.37% on LFW vs 68.82% on CALFW; VGG-Face: 97.85% on LFW vs 86.50% on CALFW; Noisy Softmax: 99.18% on LFW vs 82.52% on CALFW).
• CALFW demonstrates that attribute differences in LFW can inflate verification performance, while CALFW’s controlled setup yields a more realistic difficulty level for cross-age verification.