Source-linked AI summary
DigiFace-1M: 1 Million Digital Face Images for Face Recognition
Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, Jingjing Shen
TL;DR
Large-scale web-crawled face datasets provide recognition data but suffer from consent, label-noise, and bias concerns. This paper introduces a controllable graphics-rendered synthetic dataset and studies attribute variation and augmentation. It reports substantially better performance than GAN-generated synthetic faces and comparable accuracy after fine-tuning with a small consent-compatible real dataset.
Problem
Web-crawled face datasets used for face recognition raise privacy and consent concerns and contain label noise and bias.
Method
The paper renders a large synthetic face-recognition dataset with controllable identity, pose, appearance, accessory, camera, and environment variation.
Results
52.5% lower LFW error than SynFace improves accuracy from 91.93% to 96.17%; fine-tuning with 120K real images reaches 99.33% LFW accuracy and 93.61% across five benchmarks.
Takeaways & Limitations
Synthetic faces can support face recognition while reducing reliance on privacy-sensitive web-crawled data, especially when combined with a small number of consent-compatible real images.
Takeaways & Limitations
The synthetic pipeline cannot generate the same person at different ages, leaving an age-related domain gap and larger performance gaps on AgeDB and CALFW.
Abstract
from arXiv · showhide
State-of-the-art face recognition models show impressive accuracy, achieving over 99.8% on Labeled Faces in the Wild (LFW) dataset. Such models are trained on large-scale datasets that contain millions of real human face images collected from the internet. Web-crawled face images are severely biased (in terms of race, lighting, make-up, etc) and often contain label noise. More importantly, the face images are collected without explicit consent, raising ethical concerns. To avoid such problems, we introduce a large-scale synthetic dataset for face recognition, obtained by rendering digital faces using a computer graphics pipeline. We first demonstrate that aggressive data augmentation can significantly reduce the synthetic-to-real domain gap. Having full control over the rendering pipeline, we also study how each attribute (e.g., variation in facial pose, accessories and textures) affects the accuracy. Compared to SynFace, a recent method trained on GAN-generated synthetic faces, we reduce the error rate on LFW by 52.5% (accuracy from 91.93% to 96.17%). By fine-tuning the network on a smaller number of real face images that could reasonably be obtained with consent, we achieve accuracy that is comparable to the methods trained on millions of real face images.
1. Introduction
Existing web-crawled face datasets offer scale but raise ethical, labeling, and bias concerns. DigiFace-1M addresses these issues with controllable synthetic rendering and reports strong recognition results, including gains over GAN-generated faces and comparable accuracy after limited real-data fine-tuning.
- Motivation: Web-crawled face datasets are criticized for privacy violations, lack of informed consent, label errors, and racial, lighting, and make-up biases.The cited examples include mislabeled identities and imbalanced racial representation.
- Method: The graphics pipeline randomizes identity geometry and texture plus pose, expression, hair, accessories, cameras, and environments to encourage robust embeddings.The pipeline provides control over attribute variability and dataset scale.
- Contribution: The dataset contains 1.22M images of 110K identities and is presented as the largest public synthetic face-recognition dataset to the authors’ knowledge.It is designed to avoid privacy violations and lack of consent.
- Results: 52.5% lower LFW error than SynFace raises accuracy from 91.93% to 96.17%, while five-benchmark average error falls 46.0% from 74.75% to 86.37%.These comparisons are reported for the proposed synthetic dataset against SynFace.
- Results: Fine-tuning with only 120K real face images, 2% of MS1MV2, achieves 99.33% LFW accuracy and 93.61% average accuracy across five benchmarks.The reported accuracies are comparable to methods trained on millions of real face images.
2. Related Work
Related work contrasts real-image datasets and GAN-generated faces with a controllable synthetic alternative. The paper emphasizes varied accessory and hair rendering while preserving identity-relevant cues.
- Accessory variation: Each row in Figure 2 keeps identity fixed while varying clothing, glasses, make-up, face-wear, head-wear, and hair properties.Hair style changes only when an accessory conflicts with the original style.
- Hair variation: Figure 3 avoids randomizing hair style because most people maintain similar styles, varying hair color, density, thickness, and horizontal orientation instead.The design preserves hair style as an identity cue while still adding appearance variation.
3. Digital Faces for Face Recognition
The dataset renders photo-realistic digital faces by varying identity-defining attributes, accessories, viewpoints, environments, and image appearance. Ground-truth alignment and aggressive augmentation are designed to improve embedding robustness and narrow the synthetic-to-real gap, while the pipeline has explicit coverage limitations.
- 3.1. Face Rendering: Identity is defined by facial geometry, texture, eye color, and hair style, while other parameters vary across rendered images.Hair color, density, and thickness are randomized, whereas hair style changes only when head-wear compatibility requires it.
- 3.1. Face Rendering: Accessories, pose, expression, camera, lighting, and background are varied to expose each identity to diverse appearance conditions.Accessory sampling includes clothing, make-up, glasses, face-wear, head-wear, hands, and secondary faces.
- 3.1. Face Rendering: Ground-truth facial landmarks provide robust face alignment even when some landmarks are not visible.This avoids relying on pretrained landmark detectors for alignment.
- 3.1. Face Rendering: The pipeline cannot generate the same person at different ages and lacks some real-world attributes, including jewelry and tattoos.These limitations may leave the synthetic-data distribution mismatched with reality.
- 3.3. Data Augmentation: Aggressive on-the-fly augmentation simulates motion blur, distortion, noise, and appearance variation to reduce the synthetic-to-real domain gap.The pipeline applies random flipping and cropping alongside appearance and warping augmentations.
4. Experimental Setup
Experiments render synthetic faces with a fixed graphics configuration and evaluate models trained from scratch using standard face-verification benchmarks. The protocol spans five datasets with varied pose and age-related conditions.
- 4. Experimental Setup: Synthetic faces are rendered with Cycles at 256 samples per pixel, 256×256 resolution, and aligned crops resized to 112×112.Rendering the full dataset took approximately 10 days on 300 NVIDIA M60 GPUs.
- 4. Experimental Setup: Models are trained from scratch without ImageNet pretraining and evaluated on LFW, CFP-FP, CPLFW, AgeDB, and CALFW.The protocol reports face-verification accuracy across five benchmark datasets, including datasets with larger pose variation and age variation.
5. Experiments
Experiments show that aggressive augmentation and carefully composed synthetic data improve face-recognition accuracy, while synthetic pre-training followed by limited real-data fine-tuning approaches real-data state of the art.
- 5.1. Data Augmentation: Aggressive appearance and warping augmentation significantly improve accuracy across datasets, with warping especially helping large-pose-variation benchmarks.Synthetic images require stronger augmentation because they lack real-image imperfections; warping is particularly beneficial on CFP-FP and CPLFW.
- 5.2. Dataset Composition: Increasing horizontal and vertical pose variation improves accuracy especially on CFP-FP and CPLFW, whereas AgeDB and CALFW show similar accuracy across pose sampling choices.The result supports matching pose variation to benchmark characteristics.
- 5.2. Dataset Composition: Increasing the number of textures does not meaningfully improve accuracy, suggesting that geometry, environment, hair, accessories, and image quality contribute substantially to appearance variation.The texture library contains 208 consented face scans, while identities can be generated through varied facial geometry and environments.
- 5.3. Mixing with Real Faces: Synthetic pre-training followed by real-data fine-tuning performs better than dataset mixing, especially on large-pose-variation datasets.Training only on a small real set performs worse than training only on the synthetic data; the authors attribute the mixing gap partly to image-count imbalance.
- 5.4. Comparison to the State-of-the-Art: The proposed synthetic faces significantly outperform SynFace across all datasets under both synthetic-only and small-real-data comparisons.Table 4 uses the same LResNet50E-IR encoder; the comparison includes synthetic-only training and two-stage synthetic pre-training with real fine-tuning.
- 5.4. Comparison to the State-of-the-Art: The method reaches 96.17% on LFW with synthetic data alone, while fine-tuning on 120K real images makes performance comparable to MS1MV2-trained methods but leaves a larger gap on AgeDB and CALFW.The remaining gap is associated with the lack of explicit aging simulation and other synthetic-to-real domain differences.
6. Conclusion
The paper introduces a large-scale rendered synthetic face dataset and finds that it can outperform GAN-generated synthetic faces and approach real-data-trained methods with limited consented real data.
- 6. Conclusion: The study introduces a large-scale synthetic face-recognition dataset rendered with a graphics pipeline and evaluates augmentation and attribute sampling extensively.The dataset is intended to reduce reliance on privacy-sensitive, web-crawled face images.
- 6. Conclusion: Synthetic faces significantly outperform GAN-generated faces for face recognition, and limited real-image fine-tuning achieves accuracy comparable to methods trained on millions of web-crawled images.The conclusion frames the dataset as a step toward socially responsible face-recognition models.