Context-aware Training Image Synthesis for Trafﬁc Sign Recognition

Akira Sekizawa and Katsuto Nakajima

Department of Information Systems and Multimedia Design, Tokyo Denki University, Tokyo, Japan

Keywords:

Trafﬁc Sign Recognition, Object Detection, Synthetic Data, Data Augmentation.

Abstract:

In this paper, we propose a method for training trafﬁc sign detectors without using actual images of the trafﬁc

signs. The method involves using training images of road scenes that were synthetically generated to train

a deep-learning based end-to-end trafﬁc sign detector (which includes detection and classiﬁcation). Conven-

tional methods for generating training data mostly focus only on producing small images of the trafﬁc sign

alone and cannot be used for generating images for training end-to-end trafﬁc sign detectors, which use im-

ages of the overall scenes as the training data. In this paper, we propose a method for synthetically generating

road scenes to use as the training data for end-to-end trafﬁc sign detectors. We also show that considering the

context information of the surroundings of the trafﬁc signs when generating scenes is effective for improving

the precision.

1 INTRODUCTION

To implement advanced driver assistance systems and

achieve fully autonomous driving, researchers are ac-

tively working on object recognition technology for

recognizing humans and the objects surrounding ve-

hicles. Sensors that are used for object recognition

include RGB cameras, millimeter-wave radar, and Li-

DAR (laser radar). For example, LiDAR is effective

for detecting humans with high accuracy. In contrast,

RGB cameras are necessary for recognizing trafﬁc

signs because the system must be able to read reg-

ulations printed on the surfaces of the signs. In the

ﬁeld of image recognition by RGB cameras, end-to-

end object recognition methods based on deep learn-

ing approaches have made it possible to recognize ob-

jects with high precision and speed. However, it is

necessary to collect large amounts of diverse training

data when using end-to-end object recognition meth-

ods compared with using conventional methods.

Methods for collecting training data can be classi-

ﬁed into the following three categories: methods that

use publicly available datasets, methods that involve

manually collecting new data, and methods that in-

volve synthesizing new data. For trafﬁc sign recogni-

tion, it is possible to use published datasets in only a

few situations. Because trafﬁc sign standards are dif-

ferent in each country, trafﬁc sign datasets from one

country cannot be used for training trafﬁc sign recog-

nition systems in another country. Furthermore, man-

ual collection of trafﬁc sign images is an extremely

time-consuming process. For example, even if the

consideration is limited to only a set of regulatory

signs from all classes of trafﬁc signs in Japan, there

are still over 60 classes of signs. Furthermore, be-

cause there is bias in the locations where each class

of sign is installed, it would be extremely expensive

to travel to the locations where each trafﬁc sign is in-

stalled to photograph them. Alternatively, it is pos-

sible to collect images of the signs from the internet

instead of photographing them on-site. However, it is

difﬁcult to use the internet to collect trafﬁc sign im-

ages that are not installed in many places.

Researchers are studying methods for synthesiz-

ing data as a third method of collecting training data.

Synthetic generation of training data for trafﬁc sign

recognition is advantageous because it reduces the

cost of collecting data and makes it possible to gen-

erate trafﬁc sign datasets for any given country. A

method for generating images of single trafﬁc signs

was ﬁrst proposed by Ishida et al. in 2006 (Ishida

et al., 2006). Their method used with realistic degra-

dation involving application of various degradation

models such as blur and rotation to the template im-

ages of the trafﬁc signs. Subsequently, several other

methods have also been proposed (Hoessler et al.,

2007; Medici et al., 2008; Møgelmose et al., 2012;

Moiseev et al., 2013; Haselhoff et al., 2017). In

contrast, research on methods for generating images

of the overall scenes that include trafﬁc signs could

not be found. In several previous research projects,

researchers performed data augmentation to increase

466

Sekizawa, A. and Nakajima, K.

Context-aware Training Image Synthesis for Trafﬁc Sign Recognition.

DOI: 10.5220/0007387904660473

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 466-473

ISBN: 978-989-758-354-4

the number of images in the datasets by pasting syn-

thesized images of the trafﬁc signs on the existing

background images (Zhu et al., 2016; Peng et al.,

2017; Urˇsiˇc et al., 2017). These previous research

projects did not evaluate the precision of the trafﬁc

sign recognition model trained by only the syntheti-

cally generated trafﬁc sign images. They also did not

investigate methods for improving the trafﬁc sign im-

age synthesis method itself.

In this paper, we make two contributions. Our ﬁrst

contribution is to propose a novel method for synthet-

ically generating the training data for an end-to-end

trafﬁc sign detector. Our proposed method synthe-

sizes the trafﬁc signs at the locations where they are

supposed to be when pasting them on the background

images by focusing on the context of the surroundings

of an object, that end-to-end object detectors would

be expected to use. An example of an image of a road

scene generated using the proposed method is shown

in Figure 1. The proposed method uses only pub-

lished trafﬁc sign datasets and trafﬁc sign pictograms

to generate the training data. Therefore, it is unneces-

sary to manually collect new data. Our second contri-

bution is the evaluation of the precision of the models

that are trained using only the synthesized data. Al-

though there have been some reports on the precision

of the systems trained on a mix of synthetically gen-

erated and manually collected road scenes do exist,

studies regarding the precision when the systems are

trained on synthesized data alone are not known.

Figure 1: Road scene generated using the proposed method.

2 RELATED WORKS

2.1 Synthetic Image Generation of

Single Trafﬁc Signs

Images of single trafﬁc signs are used as the train-

ing data for pure trafﬁc sign classiﬁers. The main

method for synthetically generating images of single

trafﬁc signs is a degradation model. In this method,

degraded images are generated by applying functions

(degradation models) that express the degradation

such as a rotation, a blur, or a translation to an ideal

template image of a trafﬁc sign. A diverse set of syn-

thetic images is generated by varying the parameters

of the degradation model. Generation of training data

using degradation models was ﬁrst proposed by Baird

for application to the task of optical character recog-

nition (OCR) for optically scanned documents (Baird,

1992). Ishida et al. proposed a degradation model for

trafﬁc sign recognition with three parameters, namely,

rotation, blur, and translation (Ishida et al., 2006).

Ishida et al. also proposed a method for adaptively

determining the values of the degradation parame-

ters using a genetic algorithm (Ishida et al., 2007).

Which were manually determined based on experi-

ence. Møgelmose et al. suggested a degradation

model for trafﬁc sign recognition that included six

parameters, namely, hue change, luminosity change,

rotation, Gaussian blur, Gaussian noise, and occlu-

sion (Møgelmose et al., 2012). Moiseev et al. pro-

posed a degradation model that included hue change,

saturation change, three-dimensional rotation, projec-

tive transformation, Gaussian blur, translation, scal-

ing, and Gaussian noise (Moiseev et al., 2013). Clas-

siﬁers that were trained using the degradation model

proposed by Moiseev et al. achieved precision that

were higher than those of the classiﬁers trained using

real images in the trafﬁc sign recognition benchmark

for German trafﬁc signs (GTSRB) (Stallkamp et al.,

2012).

Another method for generating realistic synthetic

images transfers the degradation of an actual image

to a synthesized image. Haselhoff et al. proposed a

method for generating synthetic trafﬁc signs by trans-

ferring the features of the appearance of the real trafﬁc

signs to other synthetic trafﬁc signs using a Markov

random ﬁeld (Haselhoff et al., 2017). In Haselhoff’s

method, the trafﬁc sign classes of the transfer source

can be different from the classes of the destination.

Therefore, it is possible to use this method to synthet-

ically generate images of the trafﬁc signs from one

country to another country.

2.2 Synthetic Image Generation of

Overall Scenes that Include Trafﬁc

Signs

Images of the road scenes that include trafﬁc signs are

used as the training data for the end-to-end object de-

tectors. To the best of our knowledge, research that

directly investigates the synthetic generation of road

Context-aware Training Image Synthesis for Trafﬁc Sign Recognition

467

scenes for trafﬁc sign detection is not yet reported.

There are several instances of previous work in which

the researchers have generated road scenes that in-

clude trafﬁc signs as a part of a data augmentation

process.

Zhu et al. generated road scenes by pasting syn-

thetic trafﬁc signs that were generated using a degra-

dation model on background images at random lo-

cations (Zhu et al., 2016). Zhu’s degradation model

included rotation within the range from -20

to 20

scaling within the range of 20 pixels to 200 pixels,

projective transformation within the range that was

appropriate for trafﬁc signs, and random noise. Zhu

et al. combined the synthetically generated scene im-

ages with manually collected scene images and used

them for training.

Peng et al. generated scene images by extracting

the images of single trafﬁc signs from existing trafﬁc

sign datasets and pasting them on background images

at random locations (Peng et al., 2017). In Peng’s

method, actual images were used for both the traf-

ﬁc sign images and background images. Peng et al.

extracted the images of single trafﬁc signs from the

GTSRB dataset (Stallkamp et al., 2012) and used the

KITTI Road/Lane Detection Evaluation 2013 dataset

(Fritsch et al., 2013) for the background images.

Note that the researchers did not apply any additional

degradation or deformation to the trafﬁc sign images

when pasting them on background images.

Urˇsic et al. pasted both the synthetically generated

trafﬁc signsand actual trafﬁcsigns on backgroundim-

ages at random locations (Urˇsiˇc et al., 2017). Urˇsic

et al. applied deformations due to afﬁne transforma-

tions and scaling transformations to both the syntheti-

cally generated images and actual images of the trafﬁc

signs, changed their luminosity and contrast, and ap-

plied motion blur and shadow. In addition, Urˇsic et

al. proposed to avoid pasting trafﬁc sign images in

the center of the background image. The center of the

background image is normally occupied by the road,

and it is very rare for a trafﬁc sign to be located in this

part of the image.

The goal of these research projects was data aug-

mentation. The researchers trained their models by

using both collected scene images and synthetically

generated scene images simultaneously. There have

been no reports of studies in which training was per-

formed using synthetically generated scene images

alone. Furthermore, these studies did not focus on

improving the methods for synthetically generating

scene images either.

3 PROPOSED METHOD

In this section, we describe a method for syntheti-

cally generating images of scenes that include trafﬁc

signs to serve as the training data for end-to-end trafﬁc

sign detectors. Our preliminary experiments showed

that the precision of the trafﬁc sign detector that were

trained using Zhu et al’s method, in which synthet-

ically generated trafﬁc signs were pasted at random

locations (Zhu et al., 2016) (hereinafter referred to as

the random method), had a mAP that was approxi-

mately 10% lower than that of the detectors that were

trained using actual training images. An example of

a road scene that was generated using the random

method is shown in Figure 2. The procedure ﬂow is

shown in Figure 3.

Figure 2: Example of a road scene generated using the ran-

dom method.

Figure 3: Flowchart of the random method.

Although the random method exhibits a high ren-

dering quality for generating single trafﬁc signs, the

locations where the signs are pasted and sizes of the

signs are random. Therefore, the trafﬁc signs are

pasted at locations where the trafﬁc signs are not sup-

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

468

posed to exist, such as ﬂoating in the air or on center

of the road. Our hypothesis for the reason why im-

ages generated using random method yield a low per-

formance is that the context information of the sur-

roundings of the trafﬁc sign is lost in the generated

scene images. End-to-end object detectors use the

context information from the surroundings of an ob-

ject to detect and classify objects. Therefore, using

scene images in which the context information is de-

stroyed will have a negative effect on the training.

In our proposed method, we generate training data

by pasting trafﬁc signs at locations at which the trafﬁc

signs were originally located in the scene. An exam-

ple of a generated scene using our proposed method is

displayed in Figure 4. The procedure ﬂow is depicted

in Figure 5.

Figure 4: Example of a road scene generated using the pro-

posed method.

Figure 5: Flowchart of the proposed method.

To obtain the locations at which trafﬁc signs ex-

ist within the scenes, we used published trafﬁc sign

datasets. Because these datasets included numerous

scenes that include trafﬁc signs, we can use these im-

ages as the background images and paste syntheti-

cally generated trafﬁc signs at overlapping locations

where the trafﬁc signs is originally located. This

made it possible to synthetically generate scene im-

ages that preserved the context information surround-

ing the trafﬁc signs.

3.1 Collection of Background Images

Our proposed method uses images of scenes that in-

clude trafﬁc signs as background images. It is neces-

sary for the locations and shapes of the trafﬁc signs to

be annotated earlier. It is possible to use a trafﬁc sign

dataset from a country or region that is different from

the target country or region. This is because although

different countries may drive on different sides of the

road, the locations at which the trafﬁc signs are in-

stalled and the methods for installing them are practi-

cally the same. In addition, because the trafﬁc signs

are overwritten by the synthetically generated trafﬁc

signs, the method is not affected by the differences in

the trafﬁc signs between countries.

3.2 Generation of Single Trafﬁc Signs

When generating images of single trafﬁc signs, it is

necessary to consider the following four factors:

• Shape

• Size

• Rotation

• Degradation of appearance

In terms of the shape, it is necessary to choose a

trafﬁc sign of a shape that can completely covers the

original trafﬁcsign. For example, if a triangular trafﬁc

sign is pasted on top of a trafﬁc sign that is originally

circular, part of the original trafﬁc sign will still be

visible, as shown in Figure 6. This would result in an

inappropriate context surrounding the sign.

Therefore, it is desirable for the original sign and

pasted sign to have the exact same shape. However,

some shapes, such as octagons, do not appear very

frequently, and the number of existing scene images

that include these signs is small. Therefore, it is desir-

able to allow exceptions, such as allowing octagonal

signs to be pasted on top of circular signs.

In our proposed method, we do not require the

classes of the original sign and pasted sign to match.

If we were to require the classes to match, then this

Context-aware Training Image Synthesis for Trafﬁc Sign Recognition

469

Figure 6: Example of an image in which signs of different

shapes were pasted on an image.

would increase the constraints, and it would not be

possible to generate a sufﬁcient number of diverse

scene images. In addition, because there is not much

difference between the context surrounding the signs

for the different classes of signs, there is not much

advantage in requiring the types to match. As for the

size, it is necessary to make the width and height of

the synthetically generated sign to match those of the

bounding rectangle of the original sign obtained based

on the annotations.

Regarding the rotation, it is desirable to calculate

the projective transformation matrix of the original

sign and apply the same projective transformation to

the synthetically generated sign. However, because it

is difﬁcult to calculate the projective transformation

matrix of an original sign from an image, we use the

following method. First, because trafﬁc signs are po-

sitioned perpendicular to the ground, we assume that

there is no rotation in the direction of the x-axis or

z-axis. To determine the rotation in the y-axis direc-

tion, we assume that it is possible to approximate the

deformation due to rotation in the y-axis direction by

aligning the widths and heights of the bounding boxes

of the original sign and pasted sign. An example of

a synthetic sign aligning only the widths and heights

is shown in Figure 7. The projective deformation of

the trafﬁc signs that are far away and small can be

ignored. Therefore, it is possible to sufﬁciently deal

with the signs that have undergone projective trans-

formation simply by aligning the widths and heights

of the bounding rectangles, as we will demonstrate in

our evaluation later.

We processed the appearance of the signs to make

them resemble the actual images by using a degra-

dation model. Because our purpose is to paste the

synthetically generated signs within the scene, it is

desirable to consider whether the image matches

the brightness and sharpness of the background and

whether it looks realistic when determining the degra-

(a) Actual sign (b) Synthesized sign

Figure 7: Example of a synthetic image created by pasting

on top of a sign that has undergone a projective transforma-

tion.

dation to apply to the sign image. However, this topic

is left for future research. In this paper, random degra-

dation is applied.

4 EVALUATION

To evaluate the effect of road scene generation by con-

sidering the context, we trained an end-to-end object

detector,Faster R-CNN, using the followingthree sets

of training data:

1. Actual images

2. Synthesized images (proposed method)

3. Synthesized images (Zhu’s random method (Zhu

et al., 2016))

In this experiment, we used the Tsinghua–Tencent

100K trafﬁc sign dataset (TT100K) (Zhu et al.,

2016). TT100K includes 16,811 road scene images of

2048×2048 pixels with annotations, each captured in

China. The dataset includes 6,103 images in the train-

ing set; 3,067 images in the test set; and 7,641 other

images (of which 6,544 are background images that

do not include trafﬁc signs). In this experiment, we

use all images in the training set and test set. TT100K

includes total 182 classes of trafﬁc signs. In addi-

tion, TT100K also includes template images for the

128 classes. In this experiment, we use 79 of classes.

A list of the trafﬁc signs used in this experiment is

shown in Figure 8. This set of 79 classes is the inter-

section of the following two sets of classes.

• 151 trafﬁc sign classes included in the training set

• 128 trafﬁc sign classes for which a template image

has been provided

4.1 Generation of Training Data

In this experiment, we prepared three types of training

data, including actual images (Set 1), synthetic im-

ages that were generated using the proposed method

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

470

Figure 8: List of the trafﬁc signs that were used in this ex-

periment.

(Set 2), and synthetic images that were generated us-

ing Zhu’s random method (Set 3).

For Set 1, we used 6,103 images as is from the

TT100K training set.

For Set 2, we used 6,103 synthetic images that

were generated using the proposed method. We used

6,103 images from the TT100K training set as the

background images that included the trafﬁc signs

which were needed by the proposed method. We used

Moiseev’s degradation model for generating the traf-

ﬁc signs. In this experiment, we used only the bright-

ness change, saturation change, Gaussian blur, Gaus-

sian noise, and scaling from Moiseev’s degradation

model. We did not apply any geometric deforma-

tion to the sign, such as a three-dimensional rotation,

translation, or projective transformation. The param-

eters that were given to the degradation model were

tuned by hand.

For Set 3, we used 6,103 synthesized scene images

that were generated by pasting the synthesized signs

at random locations. To validate the impact of context

information on model precision alone, we set all the

conditions to be the same between Set 2 and Set 3, ex-

cept for the locations at which the signs were pasted.

Thus, we used same background images for Set 2 and

Set 3 and made the number, size, and shapes of the

synthesized signs included in the scene images to be

the same.

There is one issue that must be resolved. The

scene images used for Set 2 originally contain traf-

ﬁc signs. If these images are used as the background

images for Set 3, then the scene images would simul-

taneously include both the trafﬁc signs that were orig-

inally included in the scene images and synthetic traf-

ﬁc signs that were pasted at random locations. There-

fore, we used the inpainting method of Telea et al.

(Telea, 2004) to remove the trafﬁc signs that were

originally contained in the background images. We

set the radius parameter for the inpainting to a value

of 3. An example of trafﬁc sign removal using in-

painting is shown in Figure 9.

(a) Original image

(b) Inpainted image

Figure 9: Example of the removal of trafﬁc signs using in-

painting.

4.2 Training Method

Here, we explain the training method for faster R-

CNN. The training parameters below were selected

based on the research of Cheng et al. (Cheng et al.,

2018) who performed training using a combination

of TT100K and faster R-CNN. We used the momen-

tum SGD for the network optimization method. The

learning rate was initialized as 0.001 and was set as

0.0001 after the seventh epoch. The input resolution

for the faster R-CNN was set as 1280×1280 pixels.

Although it would have been ideal to perform the

evaluation using the original resolution of TT100K,

which was 2048×2048 pixels, we reduced the input

resolution in this evaluation to improve the training

speed. The training was conducted until the 15th

epoch. GTX 1080 Ti was used for the training. 15

epochs of the training required approximately half a

day.

4.3 Evaluation Results and Discussion

We used actual images (Set 1), synthesized images

generated using the proposed image (Set 2), and

synthesized images generated using Zhu’s random

method (Set 3) to train a total of three models. The

progress of mean Average Precision (mAP) of the

models is shown in Figure 10. The baseline shown

in the graph refers to the random method. The com-

parison of mAP in the 15th epoch after the training

has completed shows that the model trained using the

proposed method has mAP that is approximately 8%

Context-aware Training Image Synthesis for Trafﬁc Sign Recognition

471

higher than that of the model trained using the ran-

dom method. The difference between the proposed

method and random method lies in only the locations

at which the synthesized signs are pasted. Therefore,

the results demonstrate that preserving the context in-

formation is important for improving the precision

when generating the training data for an end-to-end

trafﬁc sign detector and classiﬁer.

Figure 10: Relationship between epoch and mAP of faster

R-CNN.

In contrast, mAP of the model that was trained

using the proposed method was approximately 7%

lower than that of the model trained using the actual

images. It is thought that there are two reasons for

this difference. The ﬁrst reason is that in this study,

we did not consider the context information when de-

termining the values of the parameters given to the

degradation model when generating the sign images.

For example, we did not set the brightness parameter

to match the brightness of the area surrounding the

signs. Therefore, there are cases in which a dark syn-

thesized sign is pasted on a bright area.

The second reason is that the edge between the

synthesized signs and background image is obvious.

It is necessary to use techniques such as alpha blend-

ing when pasting the trafﬁc signs on the background

images.

To analyze the difference in the performances

of the proposed model and the model trained using

the actual images, we calculated the accuracy-recall

curves for each trafﬁc sign size (Figure 11). The

graphs demonstrate two points. For large signs that

have a size in the range from 64 to 192 pixels, the

difference between the proposed method and random

method is signiﬁcant. The graphs demonstrate the ef-

fectiveness of the proposed method. However, for the

signs that have a size in the range from 32 to 64 pix-

els, the precisions of the proposed method and ran-

dom method are both lower than that of the models

trained using actual images.

Figure 11: Accuracy-recall curves based on trafﬁc sign size.

This suggests that detecting small target objects

are relatively more inﬂuenced by their surrounding

scenes. In this study, the context information was con-

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

472

sidered only in relation to selecting the locations for

pasting the road signs. In the future, we will also con-

sider the context in relation to setting the parameters

to assign to the degradation model when generating

the signs.

5 CONCLUSION

In this paper, we propose a method for training end-

to-end trafﬁc sign detectors without using actual im-

ages of the trafﬁc signs. Our proposed method en-

ables generating scene images that preserve the con-

text information surrounding the trafﬁc signs. The

proposed method achieves mAP that is approximately

8% higher than that of the conventional method, in

which the signs are pasted at random locations. This

result demonstrates that training using scene images

that preserve the context information is effective for

improving the precision. However, mAP of the pro-

posed method is approximately 7% lower than that of

the sign detectors that are trained using actual images.

The difference in the precision is high for signs that

are relatively small in the scenes compared with the

models trained using actual images. It would be pos-

sible to improve the precision by considering the con-

text information when determining the values of the

degradation parameters for generating synthetic traf-

ﬁc signs.

REFERENCES

Baird, H. S. (1992). Document Image Defect Models, pages

546–556.

Cheng, P., Liu, W., Zhang, Y., and Ma, H. (2018). Loco:

Local context based faster r-cnn for small trafﬁc sign

detection. In MultiMedia Modeling, pages 329–341.

Fritsch, J., Kuehnl, T., and Geiger, A. (2013). A new per-

formance measure and evaluation benchmark for road

detection algorithms. In International Conference on

Intelligent Transportation Systems (ITSC).

Haselhoff, A., Nunn, C., M¨uller, D., Meuter, M., and Roese-

Koerner, L. (2017). Markov random ﬁeld for image

synthesis with an application to trafﬁc sign recogni-

tion. In 2017 IEEE Intelligent Vehicles Symposium

(IV), pages 1407–1412.

Hoessler, H., W¨ohler, C., Lindner, F., and Kreßel, U.

(2007). Classiﬁer training based on synthetically gen-

erated samples. In 5th International Conference on

Computer Vision Systems (ICVS).

Ishida, H., Takahashi, T., Ide, I., Mekada, Y., and Murase,

H. (2006). Identiﬁcation of degraded trafﬁc sign sym-

bols by a generative learning method. In 18th Inter-

national Conference on Pattern Recognition (ICPR),

volume 1, pages 531–534.

Ishida, H., Takahashi, T., Ide, I., Mekada, Y., and Murase,

H. (2007). Generation of training data by degrada-

tion models for trafﬁc sign symbol recognition. IEICE

TRANSACTIONS on Information and Systems, E90-

D(8):1134–1141.

Medici, P., Carafﬁ, C., Cardarelli, E., Porta, P. P., and Ghi-

sio, G. (2008). Real time road signs classiﬁcation.

In 2008 IEEE International Conference on Vehicular

Electronics and Safety (ICVES), pages 253–258.

Møgelmose, A., Trivedi, M. M., and Moeslund, T. B.

(2012). Learning to detect trafﬁc signs: Compara-

tive evaluation of synthetic and real-world datasets. In

21st International Conference on Pattern Recognition

(ICPR), pages 3452–3455.

Moiseev, B., Konev, A., Chigorin, A., and Konushin,

A. (2013). Evaluation of trafﬁc sign recognition

methods trained on synthetically generated data. In

Advanced Concepts for Intelligent Vision Systems

(ACIVS), pages 576–583.

Peng, E., Chen, F., and Song, X. (2017). Trafﬁc sign detec-

tion with convolutional neural networks. In Cognitive

Systems and Signal Processing (ICCSIP), pages 214–

224.

Stallkamp, J., Schlipsing, M., Salmen, J., and Igel, C.

(2012). Man vs. computer: Benchmarking machine

learning algorithms for trafﬁc sign recognition. Neu-

ral Networks, 32:323 – 332.

Telea, A. (2004). An image inpainting technique based on

the fast marching method. Journal of Graphics Tools,

9(1):23–34.

Urˇsiˇc, P., Tabernik, D., Mandeljc, R., and Skoˇcaj, D. (2017).

Towards large-scale trafﬁc sign detection and recog-

nition. In 22nd Computer Vision Winter Workshop

(CVWW).

Zhu, Z., Liang, D., Zhang, S., Huang, X., Li, B., and Hu, S.

(2016). Trafﬁc-sign detection and classiﬁcation in the

wild. In 2016 IEEE Conference on Computer Vision

and Pattern Recognition (CVPR), pages 2110–2118.

Context-aware Training Image Synthesis for Trafﬁc Sign Recognition

473