The Devil of Face Recognition is in the Noise

By Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, Chen Change Loy. 2018

215 citations recorded by the source corpus.

The growing scale of face recognition datasets empowers us to train strong convolutional networks for face recognition. While a variety of architectures and loss functions have been devised, we still have a limited understanding of the source and consequence of label noise inherent in existing datasets. We make the following contributions: (1) We contribute cleaned subsets of popular face databases, i.e., MegaFace and MS-Celeb-1M datasets, and build a new large-scale noise-controlled IMDb-Face dataset. (2) With the original datasets and cleaned subsets, we profile and analyze label noise properties of MegaFace and MS-Celeb-1M. We show that a few orders more samples are needed to achieve the same accuracy yielded by a clean subset. (3) We study the association between different types of noise, i.e., label flips and outliers, with the accuracy of face recognition models. (4) We investigate ways to improve data cleanliness, including a comprehensive user study on the influence of data labeling strategies to annotation accuracy. The IMDb-Face dataset has been released on https://github.com/fwang91/IMDb-Face.

Subfields: Computer Vision and Pattern Recognition, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Signal Processing

Citations in the corpus

Diversity in Faces (2019, methodology): “…49.2% 18.0% 82.0% CelebA [9] 58.1% 42.0% 14.2% 85.8% UTKface [32] 47.8% 52.2% 35.6% 64.4% AgeDB [33] 40.6% 59.5% 5.4% 94.6% PPB [36] 44.6% 55.4% 46.4% 53.6% IMDB-Face [24] 45.0% 55.0% 12.0% 88.0% Similarly, Table 3 shows the distribution of gender and skin color/type for eight face image data sets.”
GreedyNAS: Towards Fast One-Shot NAS With Greedy Supernet (2020, background): “By dint of automatic feature engineering, deep neural networks (DNNs) have achieved remarkable success in various computer vision tasks, such as image classification [37, 36, 41, 33, 32, 15, 39], visual generation [34, 35], image retrieval [40, 7, 8, 12] and semantic comprehension [18, 17].”
Knowledge Distillation via Route Constrained Optimization (2019, methodology): “We take two popular face recognition datasets MSCeleb-1M [5] and IMDb-Face [27] as our training set and validate our method on MegaFace.”
Data Uncertainty Learning in Face Recognition (2020, background): “Some works explore the inﬂuence of label noise [39] and how to train robust FR models in this case [17, 44, 29].”
Correlation Congruence for Knowledge Distillation (2019, methodology): “MS-Celeb-1M [9] and IMDBFace [32] are used as training datasets.”
Cross-Domain Similarity Learning for Face Recognition in Unseen Domains (2021, background): “A related study [42] shows the impact of noise in training data in face recognition.”
Discriminability Distillation in Group Representation Learning (2020, methodology): “entioned experiments, we use the MS-Celeb-1M dataset for the base model and DDNet training. To demonstrate the good generalization of our method, we also train the base model and DDNet with IMDB-Face [52] dataset. IMDb-Face is a new large-scale noise-controlled dataset for face recognition. The data”
End-to-End Reconstruction-Classification Learning for Face Forgery Detection (2022, background): “We think that using larger databases for genuine faces like [8, 13, 16] to train the reconstruction network could relieve this problem in the future.”
Unequal-Training for Deep Face Recognition With Long-Tailed Noisy Data (2019, background): “…naturally think that softmax-based loss, including classical softmax and large margin softmax, could be employed to train a base model. softmax-based loss is effective, yet quite remarkably, it could be deteriorated severely by training with contaminated dataset according to recent research [35].”
CentripetalNet: Pursuing High-Quality Keypoint Pairs for Object Detection (2020, background): “With the development of deep learning, object detection has achieved great progress [10, 9, 30, 27, 25, 21, 11, 22, 1, 17, 34, 33, 36, 20, 19, 16, 39].”
Learning Self-Consistency for Deepfake Detection (2020, background): “I2G can be potentially applied to any face image or video datasets, such as IMDb-Face [50] and YouTube Faces [56], providing a stronger pretrained model for deepfake-related research.”
Agree to Disagree: Adaptive Ensemble Knowledge Distillation in Gradient Space (2020, background): “Recently, deploying deep neural networks on edge devices has appealed much attention due to their achieved success in various tasks [32, 19, 36, 34, 13, 35, 33, 15, 38].”
Metric Compatible Training for Online Backfilling in Large-Scale Retrieval (2023, unknown): “Image retrieval models [5, 10, 21, 23] have achieved remarkable performance by adopting deep neural networks for representing images.”
Deep learning with noisy labels: exploring techniques and remedies in medical image analysis (2019, background): “However, datasets collected using such methods typically suffer from very high label noise [16], [17], thus they have limited applicability in medical imaging.”
Deep Face Recognition: A Survey (2018, background): “[169] profiled the noise distribution in existing datasets in Fig.”
Sub-center ArcFace: Boosting Face Recognition by Large-Scale Noisy Web Faces (2020, background): “Even though remarkable advances have been achieved by the margin-based soft-max methods [8,15,32,31,5,39], they all need to be trained on well-annotated clean datasets [30,5], which require intensive human efforts.”