PaperlayerView exact revision on arXiv

Source-linked AI summary

WIDER FACE: A Face Detection Benchmark

Shuo Yang, Ping Luo, Chen Change Loy, Xiaoou Tang

arXiv:1511.06523v1cs.CV

TL;DR

Face detection benchmarks have not adequately captured the scale, pose, and occlusion challenges of real-world imagery. The paper introduces WIDER FACE, evaluates representative detectors, and proposes a multi-scale cascade; the dataset provides a larger, richly annotated training and evaluation source.

  • Problem

    Existing face detection datasets are too small or limited in variation and annotation richness to assess performance under real-world conditions.

  • Method

    The paper constructs WIDER FACE from diverse event images, adds bounding-box and attribute annotations, benchmarks four detectors, and proposes a scale-specialized multi-scale two-stage cascade.

  • Results

    The multi-scale cascade improves WIDER Hard average precision by 8.5% over retrained Faceness, while existing state-of-the-art algorithms reach only around 70% AP even on the Easy subset.

  • Takeaways & Limitations

    WIDER FACE provides a challenging training and evaluation source for investigating small-scale faces, occlusion, and extreme poses.

Abstract

from arXiv · show

Face detection is one of the most studied topics in the computer vision community. Much of the progresses have been made by the availability of face detection benchmark datasets. We show that there is a gap between current face detection performance and the real world requirements. To facilitate future face detection research, we introduce the WIDER FACE dataset, which is 10 times larger than existing datasets. The dataset contains rich annotations, including occlusions, poses, event categories, and face bounding boxes. Faces in the proposed dataset are extremely challenging due to large variations in scale, pose and occlusion, as shown in Fig. 1. Furthermore, we show that WIDER FACE dataset is an effective training source for face detection. We benchmark several representative detection systems, providing an overview of state-of-the-art performance and propose a solution to deal with large scale variation. Finally, we discuss common failure cases that worth to be further investigated. Dataset can be downloaded at: mmlab.ie.cuhk.edu.hk/projects/WIDERFace

1. Introduction

Face detection remains difficult in unconstrained images because pose, scale, expression, occlusion, and lighting create substantial appearance variation. WIDER FACE addresses limitations of existing benchmarks with a much larger, richly annotated dataset and evaluates representative detectors, including a multi-scale cascade approach.

  • Face detection supports downstream tasks including face alignment, recognition, verification, and parsing, but unconstrained variation makes it difficult for computers.
  • Existing benchmarks contain relatively few faces and limited variation, making real-world performance difficult to assess and exposing weaknesses under heavy occlusion, small scale, and atypical pose.
  • WIDER FACE contains 32,203 images and 393,703 labeled faces, making it 10 times larger than the current largest face detection dataset.
  • Its annotations include occlusion, pose, and event categories, enabling more detailed analysis of detector errors and failure conditions.
  • The paper benchmarks four representative algorithms and proposes a multi-scale two-stage cascade using scale-specialized convolutional networks to address large scale variation.

2. Related Work

Prior face detection research includes cascade-based, part-based, channel-feature-based, and neural-network-based methods. Existing datasets differ substantially in size, annotation richness, and variation, motivating broader benchmarking resources.

  • Face detection methods are broadly categorized as cascade based, part based, channel feature based, and neural network based.
  • Part-based methods model faces as collections of connected parts and are described as more robust to occlusion than cascade-based methods.
  • Channel-feature methods combine gradient, integral-histogram, and color-channel features with boosting classifiers, while later studies improve detection using deep learning.
  • AFW, FDDB, and PASCAL FACE provide relatively limited image or face counts, while IJB-A offers more labeled data without occlusion and pose annotations.
  • FDDB is biased toward celebrity faces, and AFW and PASCAL FACE have limited variation in face appearance and background clutter.

3. WIDER FACE Dataset

WIDER FACE is a large, richly annotated benchmark designed to expose face-detection difficulty across scale, occlusion, pose, and event variation. It also provides standardized evaluation settings and supports training with diverse positive and negative examples.

  • Dataset construction and evaluation: WIDER FACE contains 32,203 images and 393,703 labeled faces, organized into 60 event classes with 40%/10%/50% training, validation, and testing splits.The dataset supports external-data evaluation (Scenario-Ext) and training on its own training/validation partitions (Scenario-Int).
  • Dataset construction and evaluation: Annotations cover recognizable-face bounding boxes, occlusion, pose, and event categories, while very difficult low-resolution or small faces receive an Ignore flag.Bounding boxes are intended to tightly contain the forehead, chin, and cheeks; occluded faces remain annotated with estimated occlusion scale.
  • Dataset properties: WIDER FACE is more challenging than existing face-detection benchmarks, with Easy, Medium, and Hard levels defined by EdgeBox average recall rates of 92%, 76%, and 34%.These rates are measured with 8,000 proposals per image.
  • Dataset properties: Small faces remain especially difficult: faces 10-50 pixels high stay below 30% proposal detection rate even when the number of proposals reaches 10,000.Medium and large faces exceed 90% detection rate with 8,000 proposals per image.
  • Dataset properties: Atypical poses, defined using roll or pitch above 30 degrees or yaw above 90 degrees, are much harder to detect.Event categories are ranked by detection rate at 10,000 proposals and partitioned into easy, medium, and hard groups.
  • Training and baseline: WIDER FACE serves as a training source because it labels all recognizable faces and supplies diverse event-driven backgrounds containing positive and negative examples.The paper proposes a multi-scale two-stage cascade framework to address the dataset’s broad scale range.

4. Multi-scale Detection Cascade

The paper proposes a multi-scale two-stage cascade that divides face detection by scale, generating proposals and then refining them with classification and bounding-box regression. Networks are specialized for limited scale ranges to address large appearance variation.

  • Architecture: The cascade trains separate face detectors for relatively small scale ranges, using a proposal stage followed by joint classification and bounding-box regression.The first stage uses fully convolutional networks for multi-scale proposals; the second refines candidate windows.
  • Motivation: The method is designed to address WIDER FACE’s broad scale range, where small faces have particularly low generic proposal detection rates.Faces 10–50 pixels high achieve proposal detection rates below 30%.
  • Multi-scale proposal: Four scale categories are subdivided into three subclasses, with each network trained on image patches sized to its category’s upper scale bound.For example, networks use 30×30 and 120×120 patches for different scale ranges.
  • Multi-scale proposal: The proposal networks jointly learn face classification and scale classification with a multivariate cross-entropy objective.Negative samples receive scale class −1 and do not contribute to the gradient.
  • Face detection: The second stage refines proposals through joint face classification and bounding-box regression using the same CNN structure and input size for each scale category.False positives are assigned a regression target of [−1, −1, −1, −1], while regression uses Euclidean loss.

5. Experimental Results

Experiments benchmark four detectors, test WIDER FACE as a training source, and evaluate the multi-scale cascade. Results show strong difficulty across occlusion, pose, event categories, and scale, while retraining and scale specialization improve performance.

  • 5.1. Benchmarks: Faceness generally outperforms the other baseline detectors, although performance drops substantially on medium and harder settings.On the easy subset, most methods exceed 60% AP but none surpasses 75%.
  • 5.1. Benchmarks: The best baseline performance on heavy occlusion drops to 14.4%, while atypical-pose recall remains below 20%.Part-based Faceness and DPM perform relatively better under occlusion than the other baselines.
  • 5.2. WIDER FACE as an Effective Training Source: 56 out of 60 event-category accuracies improve when Faceness and ACF are retrained on WIDER FACE.The best performance on the hard “Festival” class is no more than 46% AP.
  • 5.3. Evaluation of Multi-scale Detection Cascade: The multi-scale cascade improves AP by 8.5% on WIDER Hard compared with retrained Faceness and outperforms other baselines on WIDER Medium.All evaluated models perform comparably on WIDER Easy.

6. Conclusion

WIDER FACE is presented as a large, richly annotated dataset for training and evaluating face detectors. Its emphasis is on persistent real-world challenges including small scale, occlusion, and extreme poses.

  • Conclusion: WIDER FACE provides a large, richly annotated resource for training and evaluating face detection algorithms.The dataset is intended to support further investigation of difficult detection conditions.
  • Conclusion: Even on the easy subset, where faces are typically over 50 pixels high, existing state-of-the-art algorithms reach only around 70% AP.This result indicates substantial remaining difficulty in the benchmark setting.
  • Conclusion: The paper highlights small scale, occlusion, and extreme poses as inherent face-detection challenges common in real-world applications.Surveillance and event imagery often contains faces that are small, occluded, and in atypical poses.

7. Appendix

The appendix details a multi-scale cascade detector trained with scale-specific proposal and detection networks. It also describes bounding-box refinement, hard-negative mining, and event-difficulty categorization.

  • 7. Appendix: The multi-scale cascade uses detectors specialized for relatively small face-scale ranges, with proposal generation followed by face prediction and location refinement.Each detector first generates multi-scale proposals using a fully convolutional network, then classifies candidate windows and refines positively classified boxes.
  • 7. Appendix: Four fully convolutional proposal networks are trained for face and scale classification after grouping faces into four image-size categories and three subclasses per group.The appendix references separate structures for Networks 1–4 and summarizes their model structures in Tables 5–8.
  • 7. Appendix: The detection stage fine-tunes the corresponding proposal network and jointly performs face classification and bounding-box regression.Proposals with IoU greater than 0.5 are positive; regression predicts the assigned ground-truth box positions.
  • 7. Appendix: Training uses Euclidean loss for box regression and cross-entropy loss for face classification, with additional positive crops when positives comprise less than 10% of samples.Hard-negative mining is also applied after initial proposal-network training, followed by further tuning with hard negatives.
  • 7. Appendix: Events are ranked by detection rate for scale, occlusion, and pose, then divided into easy, medium, and hard classes.The classes correspond to ranks 41–60, 21–40, and 1–20, respectively.
Loading 1511.06523v1…