PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Face Representation from Scratch

Dong Yi, Zhen Lei, Shengcai Liao, Stan Z. Li

arXiv:1411.7923v1cs.CV

TL;DR

Face recognition research lacked publicly available large-scale training data, limiting participation and fair comparison. The paper constructs CASIA-WebFace through semi-automatic Internet collection and trains a deep CNN on it. The resulting representations achieve strong performance on LFW and comparable performance to DeepFace on YTF, while surveillance performance remains below application requirements.

  • Problem

    Private large-scale face datasets were unavailable publicly, limiting research participation and comparability between methods trained on different data.

  • Method

    The paper builds CASIA-WebFace through semi-automatic Internet crawling, face clustering, duplicate checking against LFW, and manual correction, then trains a deep CNN on it.

  • Results

    The learned representations achieve comparable performance to Facebook’s DeepFace on YouTube Faces after evaluation on LFW and YTF.

  • Takeaways & Limitations

    Making CASIA-WebFace public supports standardized LFW evaluation, reproducible research, and comparability among methods using unified training and testing data.

  • Takeaways & Limitations

    Face recognition algorithms still show a large gap to surveillance requirements, especially for DIR at FAR=1% and Rank=1.

Abstract

from arXiv · show

Pushing by big data and deep convolutional neural network (CNN), the performance of face recognition is becoming comparable to human. Using private large scale training datasets, several groups achieve very high performance on LFW, i.e., 97% to 99%. While there are many open source implementations of CNN, none of large scale face dataset is publicly available. The current situation in the field of face recognition is that data is more important than algorithm. To solve this problem, this paper proposes a semi-automatical way to collect face images from Internet and builds a large scale dataset containing about 10,000 subjects and 500,000 images, called CASIAWebFace. Based on the database, we use a 11-layer CNN to learn discriminative representation and obtain state-of-theart accuracy on LFW and YTF. The publication of CASIAWebFace will attract more research groups entering this field and accelerate the development of face recognition in the wild.

1. Introduction

Face recognition had approached human-level LFW accuracy, but progress was constrained by private training data and inconsistent evaluation. The paper addresses this gap with a public dataset, a semi-automatic collection pipeline, and a deep CNN baseline.

  • 95% to 99% accuracy on LFW marked a rapid rise toward human-level face recognition performance.
  • Private large-scale datasets limited participation and made methods difficult to compare fairly because training sets differed.
  • About 10,000 subjects and 500,000 images comprise the public CASIA-WebFace dataset for standardized LFW evaluation.
  • The dataset pipeline crawls celebrity names and photos from IMDb, clusters faces for identity annotation, checks LFW overlap by name edit distance, and manually corrects errors.
  • An 11-layer CNN combining recent architectural and loss-function components achieves superior results on LFW protocols and comparable performance to DeepFace on YTF.
  • The paper contributes a public dataset, a semi-automatic construction method, and a high-performance baseline CNN for face recognition in the wild.

2. Related Work

Related work progressed from controlled face datasets and hand-crafted or shallow representations toward unconstrained datasets and learned deep CNN representations. Public data remained smaller or noisier than major private resources, motivating CASIA-WebFace.

  • Controlled datasets such as PIE and FERET enabled high performance but models learned from them generalized poorly to new environments.
  • LFW shifted evaluation toward Internet images containing varied pose, illumination, expression, resolution, and imaging conditions.
  • CACD contains 2,000 subjects and 163,446 images but includes noise and incorrect identity labels because only a small subset was manually checked.
  • Private datasets such as Facebook’s SFC, CUHK’s CelebFaces, and MSRA’s WDRef supported strong results, but were not publicly available.
  • Earlier face representations evolved from one-layer linear models to local-feature combinations with linear models, including Eigenface, Fisherface, Gabor, and LBP methods.
  • The paper follows the shift from hand-crafted shallow filters to deep CNNs that learn parameters across layers from large-scale data.

3. CASIA-WebFace Dataset

The paper builds CASIA-WebFace by semi-automatically collecting celebrity images from IMDb, assigning identities through tag- and similarity-constrained clustering, and manually correcting errors. After filtering and removing LFW name duplicates, the cleaned dataset contains 10,575 subjects and 494,414 face images.

  • 3.1. Name and Image Collection: IMDb provides celebrity names and photos, yielding 38,423 subjects and 903,304 images before face detection and subsequent filtering.Multi-view detection retains 844,126 images and detects 1,556,352 faces, though gallery overlap makes the actual counts smaller.
  • 3.2. Face Annotation: The annotation task assigns each detected face an identity and groups faces belonging to the same person.Main photos generally seed identities, while gallery images commonly contain multiple celebrities.
  • 3.2. Face Annotation: The pipeline addresses noisy web annotations, including missed face detections and missing identity annotations, through constrained matching and manual checking.These noise types are illustrated in sample photos from celebrity pages.
  • 3.2. Face Annotation: Tag-similarity clustering combines pretrained face features, celebrity seed images, single-face images, name tags, and manual deletion of falsely grouped faces.The procedure avoids constructing an all-sample similarity matrix that would exceed normal-PC memory constraints.
  • 3.2. Face Annotation: 1043 subjects matching LFW names are removed, making CASIA-WebFace an independent training set for LFW and enabling the CASIA-WebFace+LFW benchmark.Subjects with fewer than 15 face images are also removed before duplicate checking.
  • 3.3. Dataset Statistics: The cleaned CASIA-WebFace contains 10,575 subjects and 494,414 face images, but the authors cannot guarantee that every face is correctly detected and annotated.The authors state that a small amount of misclassification may not harm training and may improve robustness.

4. Learning Deep Representation

The paper learns face representations with a deep CNN that combines a very deep small-filter architecture, low-dimensional features, and identification-plus-verification objectives. The network is trained on CASIA-WebFace using online face-pair sampling and image preprocessing designed to improve robustness.

  • 4.1. Convolutional Network: The baseline CNN combines very deep architecture, low-dimensional representation, and multiple loss functions to learn face representations from CASIA-WebFace.The design integrates components from recent successful face-recognition networks.
  • 4.1. Convolutional Network: The network uses 10 convolutional layers, 5 pooling layers, and 1 fully connected layer with 3×3 filters throughout the convolutional layers.Inputs are 100×100 grayscale images; the first four pooling layers use max pooling and the last uses average pooling.
  • 4.1. Convolutional Network: Small filters and a very deep architecture reduce parameters and enhance nonlinearity, while removing redundant fully connected layers further reduces parameter count.The paper uses multiple small filters to approximate larger filters.
  • 4.1. Convolutional Network: Pool5 supplies a 320-dimensional face representation, while Softmax identification and Contrastive verification costs jointly train the network.Pool5 feeds Contrastive cost and Fc6 feeds Softmax cost; dropout regularizes the large Fc6 layer.
  • 4.1. Convolutional Network: Face images are grayscale-normalized to 100×100 using two landmarks, and mirroring doubles the training set to improve representation robustness.The selected landmark distance is fixed at 25 pixels and is described as relatively invariant to yaw variations.
  • 4.1. Convolutional Network: Contrastive positive and negative face pairs are sampled online within each batch to reduce memory and disk consumption, leaving cross-batch pair generation for future work.The implementation uses cuda-convnet and does not cover face pairs across batches.

5. Experiments

Experiments evaluate representations learned from CASIA-WebFace on LFW under standard and BLUFR protocols and on YTF. The network performs strongly against prior methods, while BLUFR results expose a remaining surveillance gap.

  • Experimental setup: 986,912 training samples result after mirroring 493,456 detected CASIA-WebFace faces.The dataset images undergo face detection, landmarking, and alignment before augmentation.
  • 5.1.1 Standard Protocol: On LFW, the base representation reaches 96.13% versus DeepFace’s 95.92% in the unsupervised setting.PCA tuning raises accuracy slightly to 96.33%.
  • 5.1.1 Standard Protocol: 97.73% under LFW’s unrestricted protocol lets single-network scheme E exceed DeepFace’s 7-network ensemble at 97.35% and approach DeepID2’s 4-network ensemble at 97.75%.The comparison uses scheme E, defined as DR with Joint Bayes on the LFW training set.
  • 5.1.2 BLUFR Protocol: BLUFR results show the deep network significantly surpasses HD-LBP + Joint Bayes, whose verification rate is 41.66% at FAR=0.1%.BLUFR emphasizes verification and open-set identification at low false-accept rates.
  • 5.1.2 BLUFR Protocol: Table 4 results are lower than Table 3 results, especially DIR at FAR=1% and Rank=1, indicating a substantial gap for surveillance applications.The paper identifies DIR as important for face surveillance or watch-list systems.
  • 5.2. Results on YTF: On YTF, cosine matching reaches 88.00% accuracy, PCA reaches 90.60%, and Joint Bayes tuning further outperforms DeepFace slightly.The evaluation samples 15 frames per video and averages 225 frame-pair similarity scores.

6. Conclusion

The work releases CASIA-WebFace and documents a reproducible pipeline for dataset construction and 11-layer CNN face representation learning.

  • CASIA-WebFace is a public large-scale dataset collected from the Internet and designed not to overlap with LFW.It supports combining the dataset with LFW for large-scale face recognition research.
  • The dataset and LFW combination can standardize evaluation and make methods comparable through unified training and testing sets.
  • The paper describes dataset construction and face representation learning with an 11-layer convolutional network.
  • The presented pipeline is intended to let researchers train a high-performance face recognition engine.
  • Future work includes enlarging the dataset, improving annotation tools and algorithms, and training a single network toward deep ensemble performance.
Loading 1411.7923v1…