PaperlayerView exact revision on arXiv

Source-linked AI summary

DSFD: Dual Shot Face Detector

Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, Feiyue Huang

arXiv:1810.10220v3cs.CV

TL;DR

Face detection remains difficult under substantial variation in scale, pose, occlusion, expression, appearance, and illumination. DSFD combines a Feature Enhance Module, Progressive Anchor Loss, and Improved Anchor Matching in a dual-shot detector, achieving superiority over state-of-the-art methods on WIDER FACE and FDDB.

  • Problem

    Face detection remains challenging because real-world faces exhibit substantial variation in scale, pose, occlusion, expression, appearance, and illumination.

  • Method

    DSFD uses a Feature Enhance Module, progressive anchor sizes across two shots, and improved anchor matching with anchor-based data augmentation.

  • Results

    DSFD demonstrates superiority over state-of-the-art face detectors on the WIDER FACE and FDDB benchmarks.

  • Takeaways & Limitations

    The proposed dual-shot design combines complementary feature, loss, and anchor-assignment techniques for face detection.

Abstract

from arXiv · show

In this paper, we propose a novel face detection network with three novel contributions that address three key aspects of face detection, including better feature learning, progressive loss design and anchor assign based data augmentation, respectively. First, we propose a Feature Enhance Module (FEM) for enhancing the original feature maps to extend the single shot detector to dual shot detector. Second, we adopt Progressive Anchor Loss (PAL) computed by two different sets of anchors to effectively facilitate the features. Third, we use an Improved Anchor Matching (IAM) by integrating novel anchor assign strategy into data augmentation to provide better initialization for the regressor. Since these techniques are all related to the two-stream design, we name the proposed network as Dual Shot Face Detector (DSFD). Extensive experiments on popular benchmarks, WIDER FACE and FDDB, demonstrate the superiority of DSFD over the state-of-the-art face detectors.

1. Introduction

DSFD addresses persistent challenges in face detection by combining enhanced feature learning, progressive anchor losses, and improved anchor matching. Experiments on FDDB and WIDER FACE demonstrate superiority over state-of-the-art detectors.

  • CNN-based detectors still struggle with faces varying in scale, pose, occlusion, expression, appearance, and illumination.
  • Existing detectors face limitations in feature learning, loss design, and anchor matching, including ignored current-layer information and anchor-context relationships.
  • DSFD introduces a Feature Enhance Module to improve feature discriminability and robustness by combining multi-level information.
  • Progressive Anchor Loss uses smaller anchors in the first shot and larger anchors in the second shot to facilitate feature learning.
  • Improved Anchor Matching integrates anchor partitioning with anchor-based augmentation to better match anchors and ground-truth faces, initializing the regressor more effectively.
  • DSFD achieves state-of-the-art performance on the FDDB and WIDER FACE benchmarks compared with existing methods.

2. Related work

Prior work advances face detection through CNNs and multi-scale features, but existing feature pyramids overlook some current-layer and anchor-context information. DSFD addresses these gaps with feature enhancement, progressive anchor sizing, and anchor-aware augmentation.

  • Early face detectors used hand-crafted features, which were later replaced by CNN-based approaches.
  • Multi-scale architectures predict or fuse features across layers to improve detection of tiny objects.
  • DSFD’s Feature Enhance Module combines current and upper feature maps through multi-level dilated convolutions to enhance feature semantics.
  • Unlike methods keeping anchor sizes fixed across stages, DSFD adaptively selects different anchor sizes for different stages.
  • DSFD combines anchor assignment with data augmentation because prior augmentation methods ignore random sampling during augmentation.

3. Dual Shot Face Detector

DSFD uses a two-shot architecture combining enhanced feature maps, progressive anchor losses, and improved anchor matching. The design uses smaller first-shot anchors, larger second-shot anchors, and anchor-aware augmentation to improve feature learning and regressor initialization.

  • Framework: DSFD extends a VGG16-based detector with six original first-shot feature maps and six enhanced feature maps produced by FEM.The framework uses conv3_3 through conv7_2 as detection layers, then transforms the original maps into enhanced features.
  • Feature Enhance Module: FEM combines upper-layer and current-layer information through normalization, up-sampling, element-wise products, concatenation, and dilated-convolution subnetworks.The module is inspired by FPN and RFB and splits feature maps into three branches with different numbers of dilated convolutional layers.
  • Progressive Anchor Loss: Progressive Anchor Loss uses smaller anchors in the first shot and larger anchors in the second shot across different feature levels.The two shot losses are weighted and summed into a single loss; the first-shot anchor size is half the second-shot size.
  • Framework: Table 1 organizes stride size, feature-map size, anchor scale, ratio, and feature counts for six original and enhanced features across two shots.These configurations specify the anchor and feature-map design underlying the dual-shot detector.
  • Progressive Anchor Loss: The first-shot loss uses smaller anchors to supervise original high-resolution feature maps, which the authors associate with detecting and classifying smaller faces.The second-shot loss uses enhanced feature maps, while prediction uses only the second-shot output without additional computational cost.
  • Improved Anchor Matching: IAM integrates anchor assignment with data augmentation by scaling faces toward anchor sizes and using anchor-based sampling with probability 2/5.The remaining 3/5 probability uses augmentation similar to SSD; the strategy targets the mismatch between discrete anchor scales and continuous face scales.

4. Experiments

Experiments evaluate DSFD through implementation details, ablations, benchmark comparisons, speed, and qualitative results. The reported findings show benefits from improved anchor matching and feature enhancement, strong benchmark performance, and robustness to varied face appearances.

  • Implementation Details: DSFD training uses pretrained VGG/ResNet backbones, Xavier initialization for new convolutions, SGD optimization, batch size 16, and staged learning-rate decay.The learning rate is 10^-3 for 40k steps, followed by 10^-4 and 10^-5 for two 10k-step periods.
  • Implementation Details: During inference, DSFD ignores first-shot outputs, applies non-maximum suppression, and returns up to 750 high-confidence bounding boxes per image.The second shot predicts the top 5k detections before suppression with jaccard overlap 0.3.
  • Analysis on DSFD: Improved anchor matching increases the average number of matched anchors for faces at different scales from 6.4 to about 6.9.The strategy also increases the number of ground-truth faces close to anchors, addressing the mismatch between discrete anchor scales and continuous face scales.
  • Analysis on DSFD: FEM outperforms RFB on WIDER FACE, including when RFB is equipped with FPN.FEM uses multi-level FPN information and stacked dilated convolutions in a multi-branch structure, producing larger receptive fields than the compared RFB design.
  • Analysis on DSFD: Feature enhancement is described as crucial, while progressive auxiliary supervision improves performance on easy, medium, and hard faces simultaneously.The progressive anchor is used to train all 12 different-scale detection feature maps.
  • Analysis on DSFD: 0.3%, 0.1%, and 0.3% improvements are attributed to improved anchor matching across three settings, while LargeBS reaches 91.2% AP in the hard setting.The reported gains correspond to the three evaluated settings, respectively.
  • Comparisons with State-of-the-Art Methods: A Res50-based DSFD runs at 22 FPS for VGA-resolution inputs on an NVIDIA P40 GPU, and qualitative examples cover multiple face variations.The visual results include scale, pose, occlusion, blur, makeup, illumination, modality, and reflection variations; blue boxes indicate confidence above 0.8.
  • Comparisons with State-of-the-Art Methods: DSFD achieves 99.1% and 86.2% on discontinuous and continuous FDDB ROC curves when false positives equal 1,000.With additional annotations, false positives are further reduced and DSFD outperforms the other methods.

5. Conclusions

The paper concludes that DSFD combines feature enhancement, auxiliary supervision with smaller anchors, and improved anchor matching to improve face detection. Experiments on FDDB and WIDER FACE are reported to demonstrate superiority over state-of-the-art detectors.

  • Conclusions: DSFD combines a Feature Enhance Module, auxiliary supervision with smaller anchors, and improved anchor matching to strengthen features and initialize regression.The improved matching strategy seeks to match anchors and ground-truth faces as far as possible.
  • Conclusions: Comprehensive experiments on FDDB and WIDER FACE demonstrate DSFD's superiority compared with state-of-the-art face detectors.The conclusion names SRN and PyramidBox as example comparison methods.
Loading 1810.10220v3…