Source-linked AI summary
Cross-Age LFW: A Database for Studying Cross-Age Face Recognition in Unconstrained Environments
Tianyue Zheng, Weihong Deng, Jiani Hu
TL;DR
LFW performance may be overly optimistic because its positive pairs often have small age gaps and its negative pairs can differ in gender and race. The paper constructs CALFW with larger same-person age gaps and attribute-matched negative pairs, finding that it provides a new face-verification challenge and sharply lowers accuracy relative to LFW.
Problem
LFW does not adequately represent cross-age variation and can make face verification resemble attribute classification through mismatched gender, race, and age-gap distributions.
Method
The paper constructs CALFW by selecting same-identity images with large age gaps, preserving LFW identities and folds, and forming negative pairs with matched gender and race.
Results
CALFW provides a new challenge for face verification; attribute-based classification accuracy reaches 51.80% on CALFW versus 86.23% on LFW.
Takeaways & Limitations
By emphasizing age-gap intra-class variation and reducing attribute differences between pair types, CALFW supports more demanding evaluation of unconstrained face verification.
Takeaways & Limitations
The benchmark focuses on same-person age gaps and does not investigate the visual similarity of different people in negative pairs.
Abstract
from arXiv · showhide
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.
1. Introduction
LFW approaches saturated face-verification accuracy but underrepresents age variation and attribute-balanced impostor comparisons. CALFW extends LFW with cross-age positive pairs, same-gender-and-race negative pairs, and unchanged verification protocols to create a more realistic challenge.
- Motivation: LFW face verification is widely used, but deep learning methods have approached nearly saturated benchmark accuracy.Reported examples include 97.35% for a deep neural network, 99.53% for DeepID, and 99.63% for FaceNet.
- Motivation: LFW positive pairs often have small age gaps, while negative pairs often have larger gaps and differing gender or race attributes.These distributions can make verification resemble age-gap or attribute classification rather than identity verification.
- CALFW construction: CALFW deliberately selects positive pairs with apparent age gaps and negative pairs whose subjects share race and gender.The construction adds age-related intra-class variation while reducing attribute differences between positive and negative pairs.
- CALFW construction: CALFW preserves LFW identities, data size, same/different verification protocol, and fold structure for direct evaluation of face-verification methods.Researchers can apply CALFW without changing the established LFW evaluation procedure.
- Benchmark objective: CALFW is designed as a more difficult benchmark that emphasizes age-gap intra-class variation in unconstrained face verification.The accompanying figure states that aging is more apparent in CALFW despite LFW containing more pictures.
2. Related Works
Related work spans controlled and unconstrained face datasets, training corpora, and stricter large-scale protocols. CALFW retains LFW’s low-barrier verification setup while manually adding age variation and matching negative-pair attributes.
- Task definitions: Face verification determines whether two images belong to the same person, unlike identification, which matches a query against a gallery.This distinction frames the task evaluated by LFW and CALFW.
- Dataset evolution: Controlled datasets such as PIE and FERET enabled high performance, but models trained on them often generalized poorly to practical unconstrained settings.Dataset development therefore shifted toward uncontrolled environments.
- Datasets: CASIA and FaceScrub provide large web-based image collections, with CASIA commonly used for deep-network training and FaceScrub containing 107,818 images of 530 celebrities.These datasets primarily support training or broad face-recognition research.
- Benchmarks: MegaFace evaluates identification with up to one million distractors, while IJB-A targets unconstrained face detection and recognition.Their protocols address large-gallery identification and unconstrained recognition rather than preserving LFW’s exact setup.
- CALFW distinction: Existing strict protocols may include cross-age comparisons, but their age gaps arise from large-scale data rather than deliberate dataset construction.The authors distinguish CALFW by manually adding age variation while retaining LFW protocols and ease of use.
3. From LFW to Cross-Age LFW
CALFW extends LFW for cross-age face verification by adding age variation within positive pairs and matching negative pairs on gender and race. Its construction preserves LFW identities and protocols while producing a more balanced, attribute-controlled benchmark whose age-gap distributions reduce attribute-based shortcuts.
- Construction Details: CALFW keeps LFW identities and verification protocols while adding age differences within same-person pairs for more realistic face verification.The dataset is built as an extension of LFW and retains its fold identities and protocols.
- Pair Formation: Positive pairs select the largest available age gap for each individual, while negative pairs use people with the same gender and race.These rules increase age variation within positive pairs and reduce attribute differences between pair classes.
- Construction Details: The construction gathers Internet images, detects and crops faces, removes duplicates, verifies labels, aligns landmarks, estimates ages, and forms pairs.The process includes crowdsourced searching and manual quality-control steps before pair formation.
- Comparison between LFW and CALFW: 4.94 versus 14.85 years: LFW positive and negative pairs have different average age gaps, respectively, enabling age-gap shortcuts in verification.In LFW, most positive gaps are below 10 years while most negative gaps exceed 10 years.
- Comparison between LFW and CALFW: CALFW positive pairs span larger age gaps, and positive and negative age-gap distributions overlap without a clear separating boundary.This confirms added aging-process variation and reduces the usefulness of age gap as a pair-class feature.
- Comparison between LFW and CALFW: 51.80% versus 86.23%: attribute-feature verification accuracy drops from LFW to CALFW when negative pairs are matched on gender and race.The result indicates little gender, race, or age-gap difference between CALFW positive and negative pairs.
- Comparison between LFW and CALFW: CALFW uses a more balanced distribution of 2, 3, or 4 images per person and includes more people in positive pairs.This differs from LFW’s imbalanced image counts and supports broader positive-pair diversity.
4. Baseline
The baseline compares metric-learning and deep-learning face-verification methods on LFW and the more challenging CALFW. Performance generally declines on CALFW, with deep learning outperforming LBP-based metric learning.
- The comparison evaluates metric-learning and deep-learning methods on LFW and CALFW to assess CALFW difficulty.Metric-learning methods use LBP features under an image-restricted setting, while the deep-learning comparison uses convolutional neural networks.
- 4.1. Comparison on metric learning.: 82.37% mean accuracy on LFW falls to 68.82% on CALFW for the best metric-learning method.KISSME drops about 10% and SVM about 11%.
- 4.2. Comparison on deep learning approaches: 97.85% accuracy on LFW for VGG drops about 11% on CALFW.The reported result compares the same network applied directly under the two datasets.
- 4.2. Comparison on deep learning approaches: 99.18% to 82.52% is the Noisy Softmax accuracy change from LFW to CALFW.The result shows a larger decline for this deep-learning network than the reported VGG comparison.
- Deep-learning approaches perform better than metric-learning methods using LBP features on the reported comparisons.The paper attributes this comparison to the effectiveness of convolutional neural networks for describing images.
5. Discussion and Future work
CALFW is presented as a unified extension of LFW that targets age-gap variation and tests face-verification methods. Empirical baselines suggest it provides a new challenge, while similar-looking negative identities remain uninvestigated.
- CALFW extends LFW with new images for the same identities, age-gap intra-class variation, matched-gender-and-race negative pairs, and a unified verification protocol.The construction is intended to evaluate real-world face verification more difficultly and focus on age-gap variation.
- Empirical results suggest that CALFW provides a new challenge for face verification.The database is validated with a convolutional-neural-network age-estimation tool and baseline metric-learning and deep-learning results.
- The benchmark does not investigate similar-looking different people alongside age gaps in the same identity.The paper proposes combining positive-pair age gaps with negative-pair looking similarity as future work.