Anchor Free Network for Multi-Scale Face Detection

Full text

Anchor Free Network for Multi-Scale Face
Detection
Chengji Wang1, Zhiming Luo2, Sheng Lian1and Shaozi Li1
1Cognitive Science Department, Xiamen University, China
2Postdoc Center of Information and Communication Engineering, Xiamen University, China
Email: {chengji, zhiming.luo, lancerlian}@stu.xmu.edu.cn, szlig@xmu.edu.cn
Abstract—Anchor-based deep methods are the most widely
used methods for face detection and have reached the state-of-the-
art result. Compared with anchor-based methods that estimates
the bounding-box rely on some pre-defined anchor boxes, anchor-
free methods perform the localization by predicting the offsets of
a pixel inside a face to its outside boundaries whose accuracies
are much more precise. However, anchor-free methods suffer the
drawback of low recall-rate mainly because 1) only using single
scale features lead to miss detection of small faces, 2) the highly
intra-class imbalance problem among different size faces. In this
paper, to address these problems, we propose a unified anchor-
free network for detecting multi-scale faces by leveraging the local
and global contextual information of multi-layer features. We also
utilize a scale aware sampling strategy to mitigate the intra-class
imbalance issue which can adaptivity select the positive samples.
Furthermore, a revised focal loss function is adopted to deal
with the foreground/background imbalance issue. Experimental
results on two benchmark datasets demonstrate the effective of
our proposed method.
I. INTRODUCTION
Human identity recognition is the key component of intel-
ligent video analysis systems which have been widely used
in access control, surveillance systems and other security
applications. In a typical face recognition system, the first step
is to localize the face in a given image which is called as
face detection. As a special case of general object detection
problem, CNNs based methods [5], [17] had achieved grati-
fying results on face detection problem. Current CNN based
face detection methods can be categorized into two categories:
Anchor-Based Methods and Anchor-Free Methods as shown
in Fig. 1.
Anchor-Based Methods: As shown in Fig. 1(a), most
CNN based detection methods like Fast-RCNN [4], Faster-
RCNN [21], SSD [12] compute the bounding box of an object
by regressing the offsets from a predefined anchor box. At
the training phase, a smoothed-L1 loss was usually used to
measure the disagreement between the estimated offsets and
the offsets corresponded to the ground truth.
Anchor-Free Methods: Different from anchor-based meth-
ods, another series of methods directly output values corre-
sponding with the position and the size of an object from
a given point (see Fig 1b). Stand for by DenseBox [6] and
UnitBox [29]. These models learn to directly predict offsets
from bounding box vertexes to points in region of interest. The
backbones of them are fully convolution neural networks [14].
Fig. 1. A visual explanation shows the difference of anchor-based and anchor-
free methods. The red bounding box is the ground truth, the blue bounding
box is a predefined anchor, and the green lines are the offsets. (a) The anchor
based methods predict the offsets based on predefined anchor. (b) The anchor-
free methods directly estimate the offsets of a point to its outside boundaries.
With the state-of-the-art performance gained by anchor-
based object detectors, as popularized in the Faster RCNN
and SSD frameworks. Recently, most of the research attention
about face detection are based on anchor-based methods, while
anchor-free methods like DenseBox and UnitBox don’t get
very promising result on the large-scale WiderFace [27] chal-
lenge dataset. Two main drawbacks of anchor-free methods are
as follows: First, the network architectures used by anchor-free
methods don’t use context information reasonably which then
suffer the issue of detecting small objects. Second, excepting
the inter-class imbalance problem, the intra-class imbalance
problem also affects the performance of model. For example,
the number of pixels of a big face with size 100 ×100 is
sixteen times of a small face with size 25 ×25.
In this paper, we focus on above problems, leveraging
anchor-free method to detect faces in unconstrained scene.
As we known, small face detection is a challenging problem.
Hu et al. [5] proved that incorporating context information
can help to detect small faces, we address this by three
mechanisms. The first one is multi-scale features fusion. Since
the features from adjacent scales are complement with each
other, a local fusion module is used to add local context
information. Then a combination of multi-scale features is
performed to increase receptive fields of small faces and
maintain more detailed information of big faces. Secondly,
we utilize scales related templates and intersection-over-union
as metric to select positive examples within face region, which
can balance the number of positive examples between big faces
2018 24th International Conference on Pattern Recognition (ICPR)
Beijing, China, August 20-24, 2018
978-1-5386-3788-3/18/$31.00 ©2018 IEEE 1554

Fig. 2. The proposed detection system backbone on top of a feed-forward VGG architecture, combined with four major part: (a) Bottom-Up Pathway. (b)
Local Contextual Information Fusion, merged local contextual information by addition. (c) Global Feature Fusion, adjust all features to the same resolution
and merged in series. (d) Detection, based on fused feature we use two independent task related network to predict pixel level classification and bounding
box regression.
and small faces. The last one, inspired by the focal loss [11],
we propose a revised focal loss function to do hard examples
mining which can further improve performance.
To sum up, the main contributions of this paper are as
follows:
1) We proposed a multi-scale feature fusion framework for
face detection, which efficiently fuse the local and global
context information.
2) A scales aware sampling strategy is utilized to alleviate
the intra-class imbalance problem.
3) A revised focal loss function is adopted to mining hard
examples.
II. RE LATE D WORK
In this section, we give a brief introduction of previous work
on face detection. Based on the detection methodologies, we
simple group previous work as following three categories:
Sliding-Window approaches: Classical detectors apply a
classifier on a dense image grid which usually use hand-crafted
features. Viola et al. [23] utilizes Haar-Like features and
AdaBoost algorithm to train a cascade face/non-face classifier.
Some works [25], [36] improve this paradigm by using more
advanced features and classifiers, and [10], [30], [31] import
CNN as feature extractors. Besides the cascade structure, [24],
[37], [16] introduce deformable part models (DPM) [3] for
face detection and achieve remarkable performance.
Anchor-Based methods: Inherited from the progress of
generic anchor based object detection methods, Jiang et al. [8]
applies Faster R-CNN [21] in face detection and get promising
results. CMS-RCNN [35] also uses Faster R-CNN [21] in face
detection by integrating body contextual information. Hu et
al. [5] demonstrates that more anchor templates can improve
recall rate as well as context information can help detecting
small faces. Recently, [17], [33] follow the detection paradigm
of SSD [12] and achieve the state-of-the-art performance.
Qiao [19] try to predict object scales in images and Zhang [34]
try to highlight object information on feature maps.
Anchor-Free methods: DenseBox[6] utilizes a fully con-
volutional neural network to regress a 4-D distance vector
of each pixel inside a face (the offset of a pixel to the top,
bottom, left and right boundaries of its candidate bounding
box), and the L2 loss function is used for training. Since
L2 loss function actually consider each offset in the 4-D
distance vector independently, UnitBox [29] propose a new
intersection-over-union loss function to jointly train this 4-D
distance vector.
III. PROP OS ED ME TH OD
In this section, we describe each component of our proposed
method in detail. Firstly, we illustrate the overall architecture
of our proposed network. Then we describe the definition of
ground truth and the loss function used to train our network.
Finally, we present a focal loss which is used to do hard
example mining.
A. Network Architecture
As shown in Fig. 2, the proposed network is a single
unified fully convolutional network that is consist of a feature
extraction sub-network and a detection module. The feature
extraction network contains a bottom-up pathway module, a
local contextual information fusion module and a final global
feature fusion module. The detection module has two sub-
networks that are a classification sub-network and a bounding-
box regression sub-network.
Bottom-Up Pathway: The bottom-up pathway is a feed
forward backbone network, which computes a pyramid of
feature maps at several scales with a scaling ratio of 2. In
this paper, VGG-16 [22] that discarded all fully connected
layers is used as the backbone. In order to capture multi-scale
information in a single network, we select outputs from the
pooling2, pooling3, pooling4 and pooling5 layers in the VGG-
16 to construct the feature pyramid {P2, P3, P4, P5}. Notice
that, this feature pyramid has strides of {4,8,16,32}with
respect to the input image.
Local Contextual Information Fusion: Inspired by the
work of Hu et al. [5] that suitable contextual information
can improve detection result, we add several fusion blocks to
1555

Fig. 3. A building block illustrate the local contextual feature fusion scheme.
embedded the local contextual information from higher layer
into lower layer, as illustrated in Fig. 3.
Firstly, We restrict each scale’s output to have the same
feature channels by adding a 3×3convolution layer re-
spectively. Then an upsampling layer is followed to enlarge
the spatial resolution of features from Pn+1 as same as Pn.
Finally, a element-wise summation is adopted to compute the
local context fused feature map F Ln. The whole process is
as follows:
F Ln=Up(Conv(Pn+1)) + Conv(Pn)(1)
where n∈[2,3,4], “Up” is a upsampling layer, and “Conv”
is a convolution layer with 128 filters.
Global Feature Fusion: Because high layer features encode
more semantic information but with coarse spatial resolutions,
which can not detect and localize small faces precisely. We use
a global feature fusion module to fuse multi-layer features. As
shown in Fig. 2(c), we first added two upsampling layers for
the feature maps F L4and F L3, then stacked together with
F L2to get the final fused global features.
Detection: The detection module contains two parallel sub-
networks, one for classification and another for bounding box
regression. The classification sub-network adds three 3×3
convolutional layers with channels equal to 128, 128 and 1
on the top of global fused features, and the final output is
a pixel probability map indicating whether a face or not. The
regression sub-network has a similar structure while the feature
channels are 384, 384 and 4, and the final outputs are 4 offset
maps. The final output size of the network is 1/4 of original
input image.
B. Loss Function
Following the paradigm of UnitBox [29], we also utilize a
multi-task loss function to train our network. The full loss L
can be represented as:
L=λcls ·Lcls +λloc ·Lloc +λΩ·Ω(w)(2)
where Lcls and Lloc are the loss of classification task and
localization task respectively, Ω(w)denote regularization term
over the network parameters w. Different loss terms are
controlled by the hyper-parameter λcls,λloc and λΩ.
Fig. 4. An illustration of scale aware balance sampling, two faces are
annotated with red box in the same image. Green boxes are templates in
green point position. Pixels within white boxes are positive examples. The
sizes are controlled by δpos.
Classification task: The classification task can be deemed
as a down-sampled segmentation task which classify each
pixel into face and non-face. Instead of only consider all the
pixels within a face as positives, we ignored some of pixels
near the boundaries that not used for training. Then for an
image, the groundtruth for each pixel is among one of the three
categories {face, non-face, ignore}which are corresponded as
y∗
i∈ {1,0,−1}.
The binary cross-entropy loss is used as the loss function
Lcls for classification, denoted as:
Lcls =−1
NX
i∈{k|y∗
k6=−1}
(y∗
i·ln byi+ (1 −y∗
i)·ln(1 −byi)) (3)
where byiis predicted probablity for pixel i, and Nis the size
of set {k|y∗
k6=−1}.
Localization task: The coordinate of a target bounding box
can be represented by its left-top point pt={xt, yt}and right-
bottom point pb={xb, yb}. For each inside pixel (xi, yi),
we can compute the position of its bounding box by a 4-
dimensional offsets vector ˆ
ti={ˆ
dxt=|xi−xt|,ˆ
dyt=|yi−
yt|,ˆ
dxb=|xi−xb|,ˆ
dyb=|yi−yb|} which are the distances
to the four boundaries. The goal for our localization task is
to estimated this offsets as accurate as possible. In order to
achieve this goal, we choose the IoU Loss [29] as the loss
function of the localization sub-network. The localization loss
is defined as follow:
Lloc =−1
NX
i∈{k|y∗
k=1}
ln(ioui)(4)
and
ioui= IoU( ˆ
Ri, R∗
i) = |ˆ
RiTR∗
i|
|ˆ
RiSR∗
i|(5)
where ˆ
Riis the predicted box of pixel iand R∗
iis its
corresponding ground truth, Nis the size of set {k|y∗
k= 1}.
This loss is only computed on those positive examples.
1556

C. Scale aware sampling
As showed in Figure 4, the size of big face is over ten
times of the small face. In a result, the spatial size of feature
maps related to small face is less than 1/10 of big faces, that
leads a bias in the model during training. In order to alleviate
this problem, we proposed a scale aware positive example
sampling strategy by implicitly incorporate positive samples
selection methodology in anchor-based methods.
For a pixel pjinside a face region, the ground truth box
is Gj(red box in Fig. 4) and a template Tjwith the same
size of Gjcentered at pj. We utilize the IoU of Gjand Tj
as a metric, and set pixel pjas positive example if the IoU is
above a predefined threshold.
lj=(1,if IoU(Gj, Tj)> δpos
−1,otherwise (6)
In this paper, we set δpos = 0.5for big faces and δpos = 0.2
for small faces whose size are small than 25 ×25.
D. Focal Loss
The original focal loss [11] is proposed to do hard example
mining during training by adaptive adjusting the weights of
different samples based on their difficulties, but it is only
used for classification task. In this paper, we extend focal
loss to do hard example mining on both classification and
localization task. For the classification task, the weights of
pixels with higher probabilities will be decreased which mean
they are easy to be classified. For the localization task, we will
lower the weights of those pixels with higher IoU with the
groundtruth. The final focal loss for both tasks are as follows:
Classification task: we use the same formula to compute
the weights as in [11].
probi=(ˆyi,if y∗
i= 1
1−ˆyi,otherwise (7)
Lcls =−1
NX
i∈{k|y∗
k=1}
(1 −probi)2ln(probi)(8)
Localization task: we regard the IoU of a estimate bound-
ing box with its groundtruth as a probability, and set the weight
as (1 −ioui).
Lloc =−1
NX
i∈{k|y∗
k=1}
(1 −ioui) ln(ioui)(9)
IV. EXP ER IM EN TS
The experiment section is organized as following: we first
introduce the basic datasets and experiment setup, then a
model analysis is conducted to verify the effectiveness of each
component in our method, finally we compare our method with
state-of-the-art on two benchmarks.
A. Datasets and Experiment Setup
The FDDB [7] and WIDER Face [27] benchmark datasets
have been used to evaluation our method. The FDDB dataset
contains 2,845 images with 5,171 annotated faces. The
WIDER Face dataset has 32,203 images with 393,703 labeled
faces, and 50% of the images are used for testing, 40% for
training and the remaining for validation. In this paper, all the
models are trained on WIDER Face training set without any
data augmentation, we report the results on the FDDB and
WIDER Face validation set.
As shown in Fig. 2(d), the sigmoid function is used to
normalize the outputs of the classification sub-network to a
probability in the range [0,1] and the ReLU is adopted as
the activation function for the regression sub-network. During
the training phase, we only use one image per-batch with size
1024×1024 due to the limitation of GPU memory. The random
crop and padding are used to generate the training images. The
Adam optimizer [9] is utilized to train our model for 40 epochs
with an initial learning rate 1e−6, a weight decay 0.0005.λloc
is set to 1 and λcls is set to 0.01. For the testing, we only test
in single scale and use a non maximal suppression (NMS)
with a threshold of 0.45 to generate the final results.
B. Model Analysis
We analyze the effectiveness of each component in our
method by extensive experiments on the Wider Face validation
set and the FDDB dataset. The Wider Face validation set
has easy, medium and hard subsets, which can be roughly
considered as large, medium and small faces.
Baseline: In order to evaluate the feature fusion strategy of
our model, we trained a face detector with the same network
architecture of UnitBox [29] as our baseline. This architecture
doesn’t have any feature fusion which only use the features of
pooling4 to compute the face probability and the features of
pooling5 to do the bounding box regression. For this baseline,
we use the same training settings but without scale aware
balance sampling and focal loss strategy.
Ablative Setting: To examine how each component affects
the final performance, we do a ablation experiment of our
model under three different settings. (1) Feature Fusion: the
model is trained only use the architecture in Fig. 2 without
another two components. (2) Scale aware sampling: the model
is trained with the balance sampled positive examples. (3)
Focal Loss: this is our complete model, consisting the scale
aware sampling and focal loss.
The results on the Wider Face validation set are reported in
TABLE I. Compared with baseline model, the feature fusion
strategy can give significant boost of mean-average-precision
(MAP). Especially on the Hard subset of Wider Face, our
feature fusion network improve MAP over 50% which means
enhancing the low layer features by the context information
from the high layer features is essential for the detection of
small face. By adding the scale aware sampling, we improve-
ment the MAP around 0.01 on the Easy and Medium subsets
while with slightly decrease on the Hard subset. Our complete
model contains all the components can further increase the
1557

0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
recall
percision
SSH−0.931
HR−0.925
CMS−RCNN−0.899
Ours−0.874
Multitask Cascade CNN−0.848
LDCF+−0.790
Faceness−WIDER−0.713
Multiscale Cascade CNN−0.691
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
recall
percision
SSH−0.921
HR−0.910
CMS−RCNN−0.874
Ours−0.851
Multitask Cascade CNN−0.825
LDCF+−0.769
Multiscale Cascade CNN−0.664
Faceness−WIDER−0.634
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
recall
percision
SSH−0.845
HR−0.806
Ours−0.696
CMS−RCNN−0.624
Multitask Cascade CNN−0.598
LDCF+−0.522
Multiscale Cascade CNN−0.424
Faceness−WIDER−0.345
Fig. 5. Precision-Recall curves on Wider Face validation set.
Fig. 6. Compared with state-of-the-art by using discontinuous ROC curves
on FDDB dataset.
MAP for all three subsets. For the FDDB dataset, we can get
a similar performance as shown in TABLE II.
TABLE I
THE ME AN- AVERAG E- PRECISION (MAP) ON WI DE R FACE VALI DATIO N
SE T.
Methods Easy Medium Hard
Baseline 0.757 0.737 0.404
Feature Fusion 0.833 0.797 0.626
Scale Aware Sampling 0.846 0.805 0.623
Focal Loss 0.874 0.851 0.696
TABLE II
TRUE P OSI TI VE RAT E ON FDDB FO R 1000 FALS E PO SIT IV ES
Baseline Feature Fusion Scale Aware Sampling Focal Loss
0.906 0.946 0.956 0.966
C. Evaluation on benchmark
FDDB dataset: We compare our face detector by using
the FDDB evaluation protocol with the state-of-the-art meth-
ods [33], [2], [18], [28], [31], [15], [30], [32], [5], [29], [20],
[26], [13], [1], [10]. We plot the discontinuous ROC curves
in Fig. 6 with the results of other methods downloaded from
FDDB official site1. Our method outperform most of the state-
of-the-art methods which indicate that our method can be used
to detect faces in unconstrained scenes.
Wider Face: Faces in the Wider Face dataset are divided
into three levels (Easy, Medium and Hard subset) according to
the difficulties of the detection. We compare our method with
several recent state-of-the-art [5], [17], [30], [18], [27], [2],
[26] on the validation set. The Precision-Recall curves of the
three levels is plot in Fig. 5, and some examples of detection
result of our method is shown in Fig. 7.
As can be seen from Fig. 5 and Fig. 6, some of the
recent anchor-based methods have better performance than our
method. There are still lots of area we need to explored to
further improve our method.
V. CONCLUSION
In this paper, a novel face detector is proposed to deal
with the performance degeneration of detecting small faces
in common anchor-free methods. After analyzing the issue
of this phenomenon, we design a multi-scale feature fusion
framework which combine local contextual information from
a wide range of layers that it is very useful for detecting small
faces. Besides, a scale aware example sampling strategy is
used to improve the recall rate of detection which successfully
import template in anchor-free detection paradigm. For the
future work, we intend to further improve the classification
strategy of background patches, and explore the relation of
IoU Loss and templates.
VI. ACK NOWLEDGEMENTS
This work is supported by the National Nature Science
Foundation of China (No. 61572409, No. U1705286 &
No. 61571188), the Natural Science Foundation of Jiangxi
Province (20171BAB212013), Fujian Province 2011 Collab-
orative Innovation Center of TCM Health Management and
Collaborative Innovation Center of Chinese Oolong Tea Indus-
tryCollaborative Innovation Center (2011) of Fujian Province.
1http://vis-www.cs.umass.edu/fddb/results.html
1558

Fig. 7. Examples of detection result on Wider Face.
REFERENCES
[1] D. Chen, S. Ren, Y. Wei, X. Cao, and J. Sun, “Joint cascade face
detection and alignment,” in European Conference on Computer Vision.
Springer, 2014, pp. 109–122.
[2] S. S. Farfade, M. J. Saberian, and L.-J. Li, “Multi-view face detection
using deep convolutional neural networks,” in ACM Int. Conf. Multime-
dia Retrieval, 2015, pp. 643–650.
[3] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan,
“Object detection with discriminatively trained part-based models,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 32, no. 9, pp. 1627–1645,
2010.
[4] R. Girshick, “Fast R-CNN,” in Proc. ICCV, 2015, pp. 1440–1448.
[5] P. Hu and D. Ramanan, “Finding tiny faces,” in Proc. CVPR, 2017, pp.
1522–1530.
[6] L. Huang, Y. Yang, Y. Deng, and Y. Yu, “Densebox: Unifying land-
mark localization with end to end object detection,” arXiv preprint
arXiv:1509.04874, 2015.
[7] V. Jain and E. Learned-Miller, “FDDB: A benchmark for face detection
in unconstrained settings,” Technical Report UM-CS-2010-009, Univer-
sity of Massachusetts, Amherst, Tech. Rep., 2010.
[8] H. Jiang and E. Learnedmiller, “Face detection with the Faster R-CNN,”
in IEEE Int. Conf. Automatic Face Gesture Recognition, 2017, pp. 650–
657.
[9] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
arXiv preprint arXiv:1412.6980, 2014.
[10] H. Li, Z. Lin, X. Shen, J. Brandt, and G. Hua, “A convolutional neural
network cascade for face detection,” in Proc. CVPR, 2015, pp. 5325–
5334.
[11] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Doll´
ar, “Focal loss for
dense object detection,” arXiv preprint arXiv:1708.02002, 2017.
[12] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C.
Berg, “SSD: Single shot multibox detector,” in Proc. ECCV, 2016, pp.
21–37.
[13] Y. Liu, H. Li, J. Yan, F. Wei, X. Wang, and X. Tang, “Recurrent
scale approximation for object detection in cnn,” in IEEE International
Conference on Computer Vision, 2017.
[14] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
for semantic segmentation,” in Proc. CVPR, 2015, pp. 3431–3440.
[15] N. Markuˇ
s, M. Frljak, I. S. Pandˇ
zi´
c, J. Ahlberg, and R. Forchheimer,
“Object detection with pixel intensity comparisons organized in decision
trees,” arXiv preprint arXiv:1305.4537, 2013.
[16] M. Mathias, R. Benenson, M. Pedersoli, and L. Van Gool, “Face
detection without bells and whistles,” in Proc. ECCV, 2014, pp. 720–
735.
[17] M. Najibi, P. Samangouei, R. Chellappa, and L. Davis, “Ssh: Single
stage headless face detector,” in Proc. CVPR, 2017, pp. 4875–4884.
[18] E. Ohn-Bar and M. M. Trivedi, “To boost or not to boost? on the limits
of boosted trees for object detection,” in Proc. ICPR, 2016, pp. 3350–
3355.
[19] S. Qiao, W. Shen, W. Qiu, C. Liu, and A. Yuille, “Scalenet: Guiding
object proposal generation in supermarkets and beyond,” in ICCV, 2017.
[20] R. Ranjan, V. M. Patel, and R. Chellappa, “A deep pyramid deformable
part model for face detection,” in IEEE Int. Conf. Biometrics Theory,
Applications and Systems (BTAS), 2015, pp. 1–8.
[21] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-
time object detection with region proposal networks,” in Proc. NIPS,
2015, pp. 91–99.
[22] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[23] P. Viola and M. J. Jones, “Robust real-time face detection,” Int. J.
Comput. Vision, vol. 57, no. 2, pp. 137–154, 2004.
[24] J. Yan, Z. Lei, L. Wen, and S. Z. Li, “The fastest deformable part model
for object detection,” in Proc. CVPR, 2014, pp. 2497–2504.
[25] B. Yang, J. Yan, Z. Lei, and S. Z. Li, “Aggregate channel features for
multi-view face detection,” in IEEE Int. Joint Conf. Biometrics (IJCB),
2014, pp. 1–8.
[26] S. Yang, P. Luo, C.-C. Loy, and X. Tang, “From facial parts responses
to face detection: A deep learning approach,” in Proc. ICCV, 2015, pp.
3676–3684.
[27] ——, “Wider face: A face detection benchmark,” in Proc. CVPR, 2016,
pp. 5525–5533.
[28] S. Yang, Y. Xiong, C. C. Loy, and X. Tang, “Face detection
through scale-friendly deep convolutional networks,” arXiv preprint
arXiv:1706.02863, 2017.
[29] J. Yu, Y. Jiang, Z. Wang, Z. Cao, and T. Huang, “Unitbox: An advanced
object detection network,” in Proc. ACM Multimedia, 2016, pp. 516–520.
[30] K. Zhang, Z. Zhang, Z. Li, and Y. Qiao, “Joint face detection and
alignment using multitask cascaded convolutional networks,” IEEE
Signal Process. Lett., vol. 23, no. 10, pp. 1499–1503, 2016.
[31] K. Zhang, Z. Zhang, H. Wang, Z. Li, Y. Qiao, and W. Liu, “Detecting
faces using inside cascaded contextual cnn,” in Proc. CVPR, 2017, pp.
3171–3179.
[32] S. Zhang, X. Zhu, Z. Lei, H. Shi, X. Wang, and S. Z. Li, “Faceboxes:
a cpu real-time face detector with high accuracy,” arXiv preprint
arXiv:1708.05234, 2017.
[33] ——, “S3fd: Single shot scale-invariant face detector,” in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition,
2017, pp. 192–201.
[34] Z. Zhang, S. Qiao, C. Xie, W. Shen, B. Wang, and A. L. Yuille,
“Single-shot object detection with enriched semantics,” arXiv preprint
arXiv:1712.00433, 2017.
[35] C. Zhu, Y. Zheng, K. Luu, and M. Savvides, “CMS-RCNN: contextual
multi-scale region-based cnn for unconstrained face detection,” in Deep
Learning for Biometrics. Springer, 2017, pp. 57–79.
[36] Q. Zhu, M.-C. Yeh, K.-T. Cheng, and S. Avidan, “Fast human detection
using a cascade of histograms of oriented gradients,” in Proc. CVPR,
vol. 2, 2006, pp. 1491–1498.
[37] X. Zhu and D. Ramanan, “Face detection, pose estimation, and landmark
localization in the wild,” in Proc. CVPR, 2012, pp. 2879–2886.
1559