Face Attribute Detection with MobileNetV2 and NasNet-Mobile

Abstract

In this paper, we propose two simple yet effective methods to estimate facial attributes in unconstrained images. We use a straight forward and fast face alignment technique for preprocessing and estimate the face attributes using MobileNetV2 and Nasnet-Mobile, two lightweight CNN (Convolutional Neural Network) architectures. Both architectures perform similarly well in terms of accuracy and speed. A comparison with state-of-the-art methods with respect to processing time and accuracy shows that our proposed approach perform faster than the best state-of-the-art model and better than the fastest state-of-the-art model. Moreover, our approach is easy to use and capable of being deployed on mobile devices.

Full text

Face Attribute Detection with MobileNetV2 and
NasNet-Mobile
Frerk Saxen, Philipp Werner, Sebastian Handrich, Ehsan Othman, Laslo Dinges, Ayoub Al-Hamadi
Faculty of Electrical Engineering and Information Technology, Neuro-Information Technology
Otto von Guericke University
Magdeburg, Germany
Frerk.Saxen@ovgu.de
Abstract—In this paper, we propose two simple yet effective
methods to estimate facial attributes in unconstrained images.
We use a straight forward and fast face alignment technique for
preprocessing and estimate the face attributes using MobileNetV2
and Nasnet-Mobile, two lightweight CNN (Convolutional Neural
Network) architectures. Both architectures perform similarly well
in terms of accuracy and speed. A comparison with state-of-the-
art methods with respect to processing time and accuracy shows
that our proposed approach perform faster than the best state-of-
the-art model and better than the fastest state-of-the-art model.
Moreover, our approach is easy to use and capable of being
deployed on mobile devices.
Index Terms—Mobile face attribute detection, MobileNetV2,
Nasnet-Mobile
I. INTRODUCTION
Estimating human face attributes is important for sev-
eral applications (e.g. face retrieval, social media or video
surveillance [1]). The estimation however is difficult due to
vast changes in appearance and shape of different attributes,
out of plane head rotations and difficult lighting conditions
in unconstrained settings. Many applications, however, re-
quire accurate and fast solutions that perform on resource-
constrained systems, such as mobile devices. E.g. it may not
be possible to upload images to the cloud for recognition due
to privacy reasons or a lack of Internet connection.
In this work we discuss previous works on facial attribute
detection and evaluate two lightweight CNN architectures with
respect to performance and speed using a straight forward
methodology that can be implemented on mobile devices.
A. Related Work
During the last decade, multiple approaches have been pro-
posed for face attribute estimation. Similar to other domains,
deep learning approaches excelled traditional methods with the
emergence of huge datasets like CelebA [2].
LNets+ANet: Liu et al. [2] propose a combined face and
attribute detection framework (similar to R-CNN [3]). Thus
it does not require any preprocessing like face and landmark
detection. However, the pipeline takes a considerable amount
of training time but the results made a remarkable improve-
ment over the state-of-the-art. With the introduction of Faster
R-CNN [4] and YOLOv3 [5] (for object detection) many
shortcomings have been eliminated. Nevertheless, since Liu
Face and
landmark
detection
Input
Rotate and
extend face
boundary
Nasnet-
Mobile
Or
MobileNet
V2
Attributes
Arch. Eyebrows
Attractive
Black Hair
Heavy Makeup
Mouth S. O.
Narrow Eyes
No Beard
Wear. Lipstick
Young
Fig. 1: Proposed attribute estimation pipeline. The face and
landmark detection as well as the Nasnet-Mobile and Mo-
bileNetV2 implementation are available online at [7], [8], [9]
and [10], respectively. The trained models will be provided on
request.
et al. [2] nobody tried to adopt these models for face attribute
detection again.
MCNN-AUX: Hand and Chellappa [6] propose a multi-
task CNN with an auxiliary network. They suggest a straight
forward CNN architecture and manually group the attributes
to train the multi-task CNN. The architecture has about 64
million parameters but only 3convolution layer and 2fully
connected layer. The training was done within 2.5hours only.
Thus, we suspect a very fast inference time on modern GPUs.
Mid-Level: Zhong et al. [11] propose to use mid-level
representations of the pre-trained FaceNet NN.1 [12] architec-
ture. They apply multi-scale spatial pooling on intermediate
layers and classify attributes at each level using a Support
Vector Machine. For each attribute the best performing layer is
chosen. The authors report that “features from the intermediate
layers demonstrate an obvious advantage [...] for attributes
describing motions of the mouth area where the gap is almost
20%”. Although this approach is interesting, it is not trained
end-to-end.
AFFACT: G¨
unther et al. [13] uses a ResNet-50 architecture.
To facilitate an alignment free attribute detection they perform
heavy augmentation during inference by applying 162 different
transformations to the detected bounding boxes. Thus, the
input tensor of their network has 162 channels. Although they
report superior classification results, it should be noted that
performing these transformations beforehand is quite costly.
DMTL: Wang et al. [14] and DMTL+: Han et al. [1]
propose, similar to Hand and Chellappa [6], a deep multi-task
learning approach. They use a slightly modified AlexNet [15]
Saxen et al., "Face Attribute Detection with MobileNetV2 and NasNet-Mobile," International Symposium on Image and Signal Processing and
Analysis (ISPA), 2019, DOI: 10.1109/ISPA.2019.8868585.
This is the accepted manuscript. The final, published version is available on IEEEXplore.
(C) 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

and model both attribute correlation and attribute heterogeneity
in a single network. Instead of the 6 subnetworks by Hand and
Chellappa [6] they use 8 subnetworks, one holistic subnetwork
and seven local nominal subnetworks.
SPLITFACE: Mahbub et al. [16] propose a CNN architec-
ture that is designed for face attribute detection of partial
occluded faces. They create severe occlusions synthetically
on CelebA and train a CNN on local patches at facial key
points of the image to robustly detect face attributes despite
heavy occlusions. However, since we do not address occlusion
explicitly we did not include SPLITFACE in our evaluation.
B. Dataset
The availability of comprehensive and well-designed
databases is crucial for any classification problem. In this
work, we use the CelebA dataset [2]. CelebA is a large-
scale dataset of celebrity images with large pose variations and
background clutter. In total, there are 202,599 face images of
10,777 identities. The dataset is split into training, validation
and test sets, with 162,770, 19,867, and 19,962 images,
respectively. For each image, 40 face attributes were annotated.
These attributes range from general attributes like sex, age and
demographic information to specific and individual character-
istics, like e.g. face shape, lip and nose size. All these labels
are binary labels, i.e. the corresponding facial attribute is either
present or not. Table I shows the performance of different
models, including a trivial classifier, that always votes for the
majority class. Thus, the trivial classifier can show the label
distribution of the test set (if we know the majority class). E.g.,
50% of the faces are smiling, and only 2.1% are bold. For the
majority of the classes, the distribution is imbalanced. Not all
given labels are correct, though. For some attributes, a clear
and objective decision might be difficult. However, Rolnick et
al. discovered that Deep Learning techniques are highly robust
against such label noise [17].
II. ME TH OD
In this work, we evaluate the performance of two new archi-
tectures for face attribute estimation: Nasnet-Mobile [18] (see
Sec. II-B) and MobileNetV2 [19] (see Sec. II-C). Compared
to most competitors our training procedure is straight forward
(see Fig. 1): We do not modify the network architecture and
we do not perform a sophisticated alignment. We discuss
the different approaches and dataset specifics. Our proposed
methods perform faster than the best state-of-the-art model
and better than the fastest state-of-the-art model (see details
in Sec. III). Sec. II-A explains the training procedure in detail
that is the same for both architectures. In Sec. II-B and
Sec. II-C the training details are given for Nasnet-Mobile and
MobileNetV2, respectively.
A. Training
As the first step in our recognition pipeline we detect the
face with a multi-scale ResNet model [7] and estimate facial
landmarks using an ensemble of regression trees [8]. We rotate
each image to align the eyes horizontally and crop the face
(centered between the eyes) with a square bounding box of 2
times the width of the face detection. We do that to capture
details like hair and necklace. Finally, we rescale the crop to
a resolution of 256 ×256 pixels.
To augment the training data we crop the image randomly
within a range of 95% to 100% of the respective axes,
independently for the image width and height. We randomly
flip the image horizontally and randomly change the saturation
and value within the HSV color space. After augmentation the
image is rescaled to 224 ×224 and converted to RGB to meet
the desired input shape of the CNNs. Also the pixel values are
scaled to the required range between -1 and 1. We monitored
the validation set accuracy during training and applied the test
set only a single time when the training finished.
B. Nasnet-Mobile
Nasnet is a scalable CNN architecture (constructed by neu-
ral architecture search) that consists of basic building blocks
(cells) that are optimized using reinforcement learning [18].
A cell consists of only a few operations (several separable
convolutions and pooling) and is repeated multiple times
according to the required capacity of the network. The mobile
version (Nasnet-Mobile) consists of 12 cells with 5.3million
parameters and 564 million multiply-accumulates (MACs).
We perform transfer learning with the pre-trained model
(pre-trained on ImageNet [20]) from [9] with the suggested
training setup (dropout = 0.5,weight decay =4e−5,batch
norm decay = 0.9997,batch norm epsilon = 1e−3). We
start with the learning rate of 0.05 (batch size = 64) and
automatically reduce the learning rate until 5e−6. Instead of
the cosine learning decay [21] used by [9] we use an automatic
learning rate scheduler [22] that estimates the slope of the
loss and reduce the learning rate (by a factor of 0.1) when
the loss has not been improved over the last 5ktraining steps.
This reduced the training time significantly without sacrificing
performance. We trained Nasnet-Mobile for 32 hours on a
single NVidia 1080 GTX.
C. MobileNetV2
MobileNetV2 is a CNN architecture for mobile devices
proposed by Sandler et al. [19]. Its first version was also
designed for face attribute detection but trained and evaluated
on Googles inhouse dataset [23]. They introduce inverted
residuals and linear bottlenecks and achieve state-of-the-art
results balancing inference time and performance for common
benchmarks like ImageNet [20], COCO [24], and VOC [25].
Our version of MobileNetV2 has 3.47 million parameters and
300 million MACs.
We used the pre-trained model mobilenet v2 1.0 224 from
[10] with the suggested training setup (depth multiplier = 1.0).
Just like the Nasnet-Mobile training we start with the learning
rate of 0.05 (batch size = 64) and automatically reduce the
learning rate until 5e−6(each step with a factor of 0.1).
We also tried the pre-trained model with depth multiplier =
1.4(mobilenet v2 1.4 224) but it was overfitting. Training
MobileNetV2 took 13 hours on our NVidia 1080 GTX.

1081091010
MACs (number of multiply-accumulates)
91
91.2
91.4
91.6
91.8
92
92.2
92.4
92.6
92.8
93
Average Performance
Nasnet-Mobile
#params: 5.3M
AFFACT(TD)
#params: 25.6M
AFFACT(L)
#params: 25.6M
DMTL+
#params: 193M
DMTL
#params: 92M
MobileNetV2
#params: 3.8M
Fig. 2: Number of multiply-accumulates (MACs) needed to
compute an inference on a single image vs. average test set ac-
curacy in a log-linear scale. The area of each model is propor-
tional to the number of parameters. Note that we estimated the
number of MACs and parameters for DMTL+ [1], DMTL [14],
and AFFACT [13] based on their reported changes to well
known architectures.
III. EXP ER IM EN TS A ND RE SU LTS
Table I shows a comparison of a trivial classifier, 5state-
of-the-art methods and our proposed methods: MobileNetV2
and Nasnet-Mobile. The results of the cited methods are
obtained from the respected publications. We report the test set
accuracies for the CelebA dataset. The trivial classifier always
votes for the majority class (obtained from the training set and
applied to the test set). We sorted the methods by their average
accuracy.
Although LNets+ANet [2] and Mid-Level [11] are out-
performed by more recent methods, both provide interesting
solutions that might provide some insights for further research.
LNets+ANet showed that a combined face and attribute de-
tection (inspired by R-CNN) can provide excellent results.
Future research might adopt recent changes to the object
detection methods (e.g. YOLOv3 [5]). Mid-Level [11] uses
feature representations from intermediate layers and obtains
remarkable results.
MCNN-AUX [6], DMTL [14], and DMTL+ [1] use a multi-
task learning approach, where the network is split into several
subnetworks. This idea seems to provide promising results.
However, from our perspective, it is not clear why splitting the
network into several subnetworks improves the performance.
A fully connected layer can learn any linear mapping that
two separate parallel fully connected layer can learn. Future
research might want to investigate this.
While AFFACT (L) [13] uses images aligned with automat-
ically detected landmarks, AFFACT (TD) performs 162 trans-
formations of detected bounding boxes without performing
any alignment. This might improve the performance slightly
(approx. 0.5%) but comes with a huge computational burden.
Our methods perform similar to MCNN-AUX, AFFACT(L),
and DMTL, but are outperformed by DMTL+ and AF-
FACT(TD). Nevertheless, Fig. 3 shows that our methods
perform faster than DMTL+ [1] and LNets+ANet [2] regarding
DMTL+
Titan X GPU
8ms; 93%
LNets+ANet
Tesla K20
49ms; 87%
Nasnet-Mobile
1080 GTX;
2.5ms; 91.6%
MobileNetV2
1080 GTX;
1,8ms; 91,5%
86
87
88
89
90
91
92
93
94
0 2 4 6 8 10 12 14 16
Accuracy (avg)
Inference time [ms]
DMTL+ - Titan X GPU LNets+Anet - Testla K20 Nasnet-Mobile - 1080 GTX MobileNetV2 - 1080 GTX
Fig. 3: Mean accuracy vs. inference time for the proposed
and other state-of-the-art approaches that report inference time.
The results for DMTL+ and LNets+ANet are taken from Han
et al. [1].
the average inference time for a single image without align-
ment. The results for DMTL+ and LNets+ANet are from [1].
Note that each approach is evaluated on a different device,
thus this comparison is biased. However, other papers did not
provide the number of multiply-accumulates (MACs), which
would allow for a better comparison. Thus, we estimate the
number of MACs and parameters based on their reported
changes to well known architectures. DMTL+, and DMTL use
a modified AlexNet, and AFFACT uses ResNet-50 (with one
additional layer). In Fig. 2 we report the number of MACs
vs. average test set accuracy in a log-linear-scale. The area
of each circle is proportional to the number of parameters
of each model and also proportional to the required memory.
This is particularly interesting for mobile applications with
limited resources, because storing huge amounts of parameters
in memory might not be possible. Compared to DMTL+, our
MobileNetV2 model needs 2.9times less MACs, requires 56
times less memory, and performs 1.1% worse in terms of
accuracy. Although Nasnet-Mobile und MobileNetV2 perform
very similar, Nasnet-Mobile requires 1.5M more parameters
and is about 40% slower. Even though we tried several regu-
larization methods, Nasnet-Mobile was not able to utilize its
higher capacity. Usually, models with higher capacity just need
more regularization to improve performance. Nevertheless, the
test devices weren’t too different and even if all methods would
have been tested on the same device, we do not believe that
their order would change.
IV. CONCLUSION AND FUTURE WOR K
We addressed the problem of estimating facial attributes
from RGB images for mobile devices. Face attribute estimation
is important for human machine interaction (HCI) systems
by providing relevant context information like age, sex and
ethnicity of the interacting user. We evaluated the face at-
tribute detection performance on two mobile architectures:
Nasnet-Mobile and MobileNetV2. Our experimental evalua-

5 o Clock Shadow
Arched Eyebrows
Attractive
Bags Under Eyes
Bald
Bangs
Big Lips
Big Nose
Black Hair
Blond Hair
Blurry
Brown Hair
Bushy Eyebrows
Chubby
Double Chin
Eyeglasses
Goatee
Gray Hair
Heavy Makeup
High Cheekbones
Male
Trivial 90.0 71.6 49.6 79.7 97.9 84.4 67.3 78.8 72.8 86.7 94.9 82.0 87.0 94.7 95.4 93.5 95.4 96.8 59.5 51.8 61.4
LNets+ANet [2] 91 79 81 79 98 95 68 78 88 95 84 80 90 91 92 99 95 97 90 87 98
Mid-Level [11] 93.3 82.5 80.8 82.2 97.8 95.6 69.9 82.6 86.0 94.9 96.2 84.2 91.9 94.9 96.2 99.5 97.1 97.8 90.1 86.1 98.1
MCNN-AUX [6] 94.5 83.4 83.1 84.9 98.9 96.0 71.5 84.5 89.8 96.0 96.2 89.2 92.8 95.7 96.3 99.6 97.2 98.2 91.5 87.6 98.2
AFFACT (L) [13] 94.8 83.9 82.8 85.2 99.1 96.1 72.5 84.4 90.5 96.2 96.0 88.5 92.3 95.7 96.4 99.6 97.5 98.3 92.0 87.6 98.2
AFFACT (TD) [13] 94.4 85.5 81.4 84.2 99.0 95.5 84.0 83.0 91.6 95.7 96.1 85.7 92.8 95.7 96.8 99.5 96.7 98.1 91.7 88.3 98.7
DMTL [14] 94 86 83 85 99 96 79 85 91 96 96 88 92 96 97 99 97 98 92 88 98
DMTL+ [1] 95 86 85 85 99 99 96 85 91 96 96 88 92 96 97 99 99 98 92 88 98
MobileNetV2 94.9 84.2 82.7 85.6 99.0 96.2 72.2 84.6 89.9 96.0 96.3 88.8 92.8 95.8 96.5 99.6 97.5 98.3 91.8 87.7 98.4
Nasnet-Mobile 94.9 84.1 83.2 85.4 99.1 96.2 72.3 85.0 90.2 96.1 96.4 89.4 92.9 95.9 96.5 99.7 97.6 98.2 92.0 87.8 98.4
Mouth Slightly Open
Mustache
Narrow Eyes
No Beard
Oval Face
Pale Skin
Pointy Nose
Receding Hairline
Rosy Cheeks
Sideburns
Smiling
Straight Hair
Wavy Hair
Wearing Earrings
Wearing Hat
Wearing Lipstick
Wearing Necklace
Wearing Necktie
Young
Average
Trivial 50.5 96.1 85.1 85.4 70.4 95.8 71.4 91.5 92.8 95.4 50.0 79.0 63.6 79.3 95.8 47.8 86.2 93.0 75.7 79.9
LNets+ANet [2] 92 95 81 95 66 91 72 89 90 96 92 73 80 82 99 93 71 93 87 87
Mid-Level [11] 92.6 96.6 86.9 95.4 70.6 96.7 76.2 92.1 94.3 97.4 92.1 80.0 77.3 86.7 98.8 92.3 85.8 94.4 87.5 89.8
MCNN-AUX [6] 93.7 96.9 87.2 96.0 75.8 97.0 77.5 93.8 95.2 97.8 92.7 83.6 83.9 90.4 99.0 94.1 86.6 96.5 88.5 91.3
AFFACT (L) [13] 93.8 97.0 87.6 96.2 76.7 97.1 77.1 93.7 95.2 97.8 92.8 85.0 85.7 91.0 99.1 93.7 88.3 96.9 88.8 91.5
AFFACT (TD) [13] 93.9 96.4 93.8 96.0 76.8 96.8 77.6 94.9 95.2 97.3 92.9 85.5 87.9 92.0 98.9 92.7 90.3 96.8 88.7 92.0
DMTL [14] 94 97 90 96 78 97 78 94 96 98 93 85 87 91 99 93 89 97 90 92.1
DMTL+ [1] 94 97 90 97 78 97 78 94 96 98 94 85 87 91 99 93 89 97 90 92.6
MobileNetV2 94.1 97.1 87.8 96.5 76.0 96.8 77.4 93.6 95.1 97.9 93.1 84.6 85.0 90.8 99.1 93.9 87.4 96.8 88.4 91.5
Nasnet-Mobile 94.1 97.1 87.6 96.4 76.4 97.0 77.8 94.0 95.2 98.0 93.1 85.0 85.6 91.0 99.1 94.0 87.5 96.8 88.5 91.6
TABLE I: CelebA [2] test set accuracy for each individual attribute and the average accuracy across all attributes. Color indicates
rank of method from red (worst) to green (best) – for each attribute individually. Trivial refers to the trivial classifier that
always votes for the majority class. The results of the state-of-the-art approaches are obtained from the respected publications.
See Sec. III for a detailed discussion of the results.
tion showed that using architectures for resource limited appli-
cations can perform almost as good as current top performing
architectures. Our methods are fast, accurate, and easy to
implement.
We also discussed the contributions of previous works. Es-
pecially the results by LNets+ANet [2] show interesting ideas
for future research towards a complete end-to-end training e.g.
with SSD [26] or Yolo [5]. Also, works by Wang et al. [14]
and Han et al. [1] have shown that multi-task learning can
improve face attribute detection performance. They showed
that including face recognition or other face related tasks into
the same network improves face attribute detection as well.
ACK NOW LE DG EM EN T
This work has been funded by the Federal Ministry
of Education and Research (BMBF), projects 03ZZ0443G,
03ZZ0459C, and 03ZZ0470. The sole responsibility for the
content lies with the authors.
REFERENCES
[1] Hu Han, Anil K. Jain, Fang Wang, Shiguang Shan, and Xilin Chen,
“Heterogeneous Face Attribute Estimation: A Deep Multi-Task Learning
Approach,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, jun 2017.
[2] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang, “Deep
Learning Face Attributes in the Wild,” in 2015 IEEE International
Conference on Computer Vision (ICCV). dec 2015, pp. 3730–3738,
IEEE.
[3] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik, “Rich
Feature Hierarchies for Accurate Object Detection and Semantic Seg-
mentation - IEEE Conference Publication,” in Computer Vision and
Pattern Recognition (CVPR), 2014 IEEE Conference on, Columbus, OH,
USA, 2014.
[4] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, “Faster
R-CNN: Towards Real-Time Object Detection with Region Proposal
Networks,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, pp. 1137—-1149, 2016.
[5] Joseph Redmon and Ali Farhadi, “YOLOv3: An Incremental Improve-
ment,” arXiv, 2018.
[6] Emily M. Hand and Rama Chellappa, “Attributes for Improved At-
tributes: A Multi-Task Network Utilizing Implicit and Explicit Re-
lationships for Facial Attribute Classification,” in Thirty-First AAAI
Conference on Artificial Intelligence, apr 2017, pp. 4068—-4074.
[7] Davis E. King, “Easily Create High Quality Object Detectors with Deep
Learning,” 2016.
[8] Davis E. King, “Real-Time Face Pose Estimation,” 2014.
[9] Github Repository, “Slim Nasnet,” 2018.
[10] Github Repository, “Slim Mobilenet,” 2018.
[11] Yang Zhong, Josephine Sullivan, and Haibo Li, “Leveraging mid-level
deep representations for predicting face attributes in the wild,” in IEEE
International Conference on Image Processing (ICIP). sep 2016, pp.
3239–3243, IEEE.
[12] Florian Schroff, Dmitry Kalenichenko, and James Philbin, “FaceNet:

A Unified Embedding for Face Recognition and Clustering,” in 2015
IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
jun 2015, pp. 815–823, IEEE.
[13] Manuel G¨
unther, Andras Rozsa, and Terranee E. Boult, “AFFACT:
Alignment-free facial attribute classification technique,” in IEEE Inter-
national Joint Conference on Biometrics (IJCB). oct 2017, pp. 90–99,
IEEE.
[14] Fang Wang, Hu Han, Shiguang Shan, and Xilin Chen, “Deep Multi-
Task Learning for Joint Prediction of Heterogeneous Face Attributes,”
in 2017 12th IEEE International Conference on Automatic Face and
Gesture Recognition (FG 2017). may 2017, pp. 173–179, IEEE.
[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, “ImageNet
Classification with Deep Convolutional Neural Networks,” in Pro-
ceedings of the 25th International Conference on Neural Information
Processing Systems - Volume 1. 2012, pp. 1097–1105, Curran Associates
Inc.
[16] Upal Mahbub, Sayantan Sarkar, and Rama Chellappa, “Segment-based
Methods for Facial Attribute Detection from Partial Faces,” arXiv, 2018.
[17] David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit, “Deep
Learning is Robust to Massive Label Noise,” in International Conference
on Learning Representations (ICLR), may 2018.
[18] Barret Zoph, Google Brain, Vijay Vasudevan, Jonathon Shlens, and
Quoc V Le, “Learning Transferable Architectures for Scalable Image
Recognition,” arXiv, 2017.
[19] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov,
and Liang-Chieh Chen, “MobileNetV2: Inverted Residuals and Linear
Bottlenecks,” arXiv, jan 2018.
[20] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev
Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla,
Michael Bernstein, Alexander C. Berg, and Li Fei-Fei, “ImageNet Large
Scale Visual Recognition Challenge,” International Journal of Computer
Vision, vol. 115, no. 3, pp. 211–252, dec 2015.
[21] Ilya Loshchilov and Frank Hutter, “SGDR: STOCHASTIC GRADIENT
DESCENT WITH WARM RESTARTS,” in 5th International Confer-
ence on Learning Representations, Palais des Congr`
es Neptune, Toulon,
France, 2017.
[22] Davis E. King, “Automatic Learning Rate Scheduling That Really
Works,” 2018.
[23] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko,
Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam,
“MobileNets: Efficient Convolutional Neural Networks for Mobile Vi-
sion Applications,” arXiv, vol. 1704.04861, apr 2017.
[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Per-
ona, Deva Ramanan, Piotr Doll´
ar, and C. Lawrence Zitnick, “Microsoft
COCO: Common Objects in Context,” in European Conference on
Computer Vision. 2014, pp. 740–755, Springer, Cham.
[25] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I.
Williams, John Winn, and Andrew Zisserman, “The Pascal Visual Object
Classes Challenge: A Retrospective,” International Journal of Computer
Vision, vol. 111, no. 1, pp. 98–136, jan 2015.
[26] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott
Reed, Cheng-Yang Fu, and Alexander C. Berg, “SSD: Single Shot
MultiBox Detector,” in European Conference on Computer Vision
(ECCV). 2016, pp. 21–37, Springer, Cham.