Street-view Change Detection via Siamese Encoder-decoder Structured

Convolutional Neural Networks

Xinwei Zhao

, Haichang Li

, Rui Wang

, Changwen Zheng

and Song Shi

Institute of Software Chinese Academy of Sciences, University of Chinese Academy of Sciences, Beijing, China

Science and Technology on Integrated Information System Laboratory,

Institute of Software Chinese Academy of Sciences, Beijing, China

Teleware Info & Tech (Fujian) Co.,LTD, Fujian Province, China

Keywords:

Change Detection, Semantic Segmentation, Siamese Network, Deep Learning.

Abstract:

In this paper, we propose a siamese encoder-decoder structured network for street scene change detection. The

encoder-decoder structures have been successfully applied for semantic segmentation. Our work is inspired

by the similarity between change detection and semantic segmentation, and the success of siamese network

in comparing image patches. Our method is able to precisely detect changes of street scene at the presence

of irrelevant visual differences caused by different shooting conditions and weather. Moreover, the encoder

and decoder parts are decoupled. Various combinations of different encoders and decoders are evaluated in

this paper. Experiments on two street scene datasets, TSUNAMI and GSV, demonstrate that our method

outperforms previous ones by a large margin.

1 INTRODUCTION

Change detection, i.e., ﬁnding meaningful changes

from registered image pairs of the same region but

captured at different time, is an important task in com-

puter vision. Speciﬁcally, given the registered image

pairs, we need to label each pixel as positive if it

has changed at semantic level or negative otherwise,

and produce a change mask at last, as Fig. 1 shows.

Change detection has been widely applied in several

areas including sandy land monitoring, offshore oil

spill detection, and urban planning, etc.

Change detection is quite challenging as a lot of

factors introduce irrelevant visual differences to the

image pairs, such as the differences in shooting equip-

ment, shooting conditions and weather. An example

of street-view image pairs is shown in Fig. 1. In Fig.

1, the appearance of the buildings is quite different as

they are captured in different weather. This demon-

strates that objects in such image pairs are likely to

show large variability even if they are unchanged.

Change detection has been studied for decades

(Singh, 1989). Most methods are pixel-based such

as image differencing and change vector analysis. In

most cases, these methods are unable to exclude the

aforementioned irrelevant visual differences, and can-

not precisely detect the changes as desired. Howe-

ver, convolutional networks are robust to handle these

problems. In this paper, we will explore street-view

change detection using deep learning methods.

The goal of this paper is to propose a method

to detect semantic changes precisely from registe-

red street-view image pairs at the presence of the

irrelevant changes. Inspired by the development

of semantic segmentation (Long et al., 2015; Chen

et al., 2017; Chen et al., 2018), we build a sia-

mese encoder-decoder structured convolutional net-

work (SEDS-CNN) to handle the change detection

problem. The ﬂowchart of the SEDS-CNN is shown

in Fig. 2.

Speciﬁcally, the two encoders of the siamese con-

volutional networks have an identical structure and

share the same weights, and the two decoders have

the same characteristics. They are designed to extract

the semantic information of image pairs. The last part

of SEDS-CNN, the differentiator, takes the absolute

difference of two feature maps generated by the deco-

ders and produces 2-channel feature maps to denote

the probability of changes and non-changes.

Moreover, we explore different combinations of

encoders and decoders to construct the SEDS-CNN.

Experiments on two publicly available datasets TSU-

NAMI and GSV (Sakurada and Okatani, 2015) show

that the proposed SEDS-CNN model outperforms the

Zhao, X., Li, H., Wang, R., Zheng, C. and Shi, S.

Street-view Change Detection via Siamese Encoder-decoder Structured Convolutional Neural Networks.

DOI: 10.5220/0007407905250532

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 525-532

ISBN: 978-989-758-354-4

525

Figure 1: One sample from TSUNAMI dataset (Sakurada and Okatani, 2015). The top and middle images are the same place

captured at different time, and the bottom image is the corresponding change mask. Black blocks are changed regions while

white ones are not changed. Note that although the two images on the top show large variability because of the sunlight, most

regions including the buildings on the left and right side are not regarded as changes.

existing approaches by a large margin.

The main contributions of this paper are listed as

follows:

1. A siamese network, called SEDS-CNN, is propo-

sed for street-view change detection. This net-

work is an end-to-end framework, which can ex-

tract semantic information and predict changes at

semantic level.

2. Experiments on two typical street-view datasets

demonstrate the effectiveness of the proposed

method, showing our model’s robustness to irrele-

vant visual differences. Our method outperforms

the previous ones by a large margin.

The remainder of our paper is organized as fol-

lows: Section 2 provides an overview of related

work on change detection and semantic segmenta-

tion. Section 3 presents the structure of our networks.

Section 4 details the experiment conﬁguration and

analyzes the results. Section 5 concludes our work.

2 RELATED WORK

2.1 Change Detection

Change Detection has been studied for several deca-

des (Singh, 1989; Radke et al., 2005; Tewkesbury

et al., 2015) and witnessed great success in many

areas especially remote sensing. Among these con-

ventional methods, the pixel and post-classiﬁcation

methods still remain popular even they have been

proposed for nearly 30 years. While these algo-

rithms produce good results on remote sensing ima-

ges, their robustness is far from enough to over-

come the irrelevant visual differences in optical image

pairs, whose ﬁnal appearances are much more ea-

sily affected by uncontrollable factors such as chan-

ging weather and seasons, or small camera displace-

ment. A number of other methods focus on proba-

bility graph model including Markov Random Field

(Bruzzone and Prieto, 2000), Conditional Random

Field (Li et al., 2018), and Restricted Boltzmann ma-

chine (Gong et al., 2016).

In recent years, deep convolutional neural net-

works have shown striking power in computer vision.

Since then, more sophisticated CNNs are proposed

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

526

to tackle several classic vision problems. SSD (Liu

et al., 2016), YOLO (Redmon and Farhadi, 2017) and

FPN (Lin et al., 2017) are proposed to solve object

detection problems. FCN (Long et al., 2015), Dee-

plab(Chen et al., 2018) and many other fully con-

volutional based methods are designed to deal with

semantic segmentation. All these networks perform

far better than those methods that don’t use deep lear-

ning. However, few deep learning models have been

proposed to deal with change detection, at least for

supervised methods.

(Zagoruyko and Komodakis, 2015) proposed a si-

amese CNN architecture to compare image patches to

identify if objects in two images captured from diffe-

rent angles are the same one. Although it is not im-

possible to compare the entire image pairs in a patch-

by-patch way, it’s nearly unpractical both in terms of

complexity and performance. (Sakurada and Oka-

tani, 2015) used a VGG16 network fully pretrained

for large-scale object recognition task to help extract

features, and put them together with some manually

designed features to detect changes in street scene

image pairs. (Alcantarilla et al., 2018) proposed a de-

convolutional neural network to perform change de-

tection in street-view images. They superimposed one

image on top of the other to obtain a 6-channel image,

and then feed it to a simple deconvotlutional network,

which comprises 4 convolutional and 4 deconvolutio-

nal layers. Although deep learning are used in these

models, successful techniques in semantic segmenta-

tion such as dilated convolution and multi-scale py-

ramid pooling(Zhao et al., 2017; Chen et al., 2018)

are not applied. Besides, they didn’t merge low-level

feature maps either, which have rich boundary infor-

mation.

2.2 Semantic Segmentation

Deep convolutional networks have been successfully

applied both in recognition and semantic segmenta-

tion. (Long et al., 2015) was the ﬁrst to propose

the fully convolutional neural network (FCN) trained

end-to-end to solve the dense pixel-wise prediction

tasks. However, the loss of spatial information cau-

sed by pooling layers is the major reason to restrict

its performance. In order to tackle this problem, se-

veral techniques are proposed to preserve spatial in-

formation as the network goes deeper. (Yu and Kol-

tun, 2015) proposed dilated convolution to replace

pooling and convolutional layers at the latter part of

convolutional networks, and it indeed expands the re-

ceptive ﬁelds and preserves the resolution of feature

maps at the same time, without increasing the num-

ber of parameters. (Ronneberger et al., 2015) suggests

concatenating low-level features to high-level ones to

compensate for the loss of spatial information. (Chen

et al., 2017) proposed to use global pooling and ASPP

(Atrous Spatial Pooling Pyramid) to capture multi-

scale information. Based on that, (Chen et al., 2018)

merged low-level feature maps to ASPP, and obtained

the state-of-art semantic segmentation model evalua-

ted on PASCAL VOC dataset.

Most proposed networks for semantic segmenta-

tion can be explained from an encoder-decoder per-

spective, in which encoders are used to extract spatial

and semantic information while decoders are used to

gather them to give each pixel a semantic label. This

is quite similar to change detection, where each pixel

is labeled changed or unchanged. In this case, we can

regard changes or non-changes as a kind of seman-

tic label, and this is the motivation behind our appro-

ach: use encoder-decoder structures as the backbone

for our model.

3 PROPOSED MODEL

In this section, we present the proposed SEDS-CNN

model. The overall ﬂowchart is shown in Fig. 2.

From Fig. 2, we can ﬁnd that the network has three

parts: encoder, decoder, and differentiator. The struc-

tures of the ﬁrst two parts are identical and share the

same parameters. This is because (1) the two images

from each pair are unordered and we can not specify

which image precedes the other one, and (2) two ima-

ges of each pair should be projected to the same se-

mantic feature space to be compared. The extracted

semantic features produced by decoders are then em-

ployed by the differentiator and the following com-

ponents for change detection. We will describe the

above three parts in detail in the following paragraphs.

3.1 Encoder

The appearance for the same object could be varia-

ble in different images, even if they are unchanged at

semantic level. Inspired by this point, we intend to

project the original RGB image into the semantic fe-

ature space, which is favorable for change detection.

The above idea can be achieved by the encoding step.

The encoder part of SEDS-CNN model consists

of multiple convolutional layers, dilated convolutio-

nal layers and max pooling layers. Their functions are

merely to generate semantic features from the ima-

ges. As the input images go through these layers, the

extracted features become more and more abstract,

which have more semantic information.

Street-view Change Detection via Siamese Encoder-decoder Structured Convolutional Neural Networks

527

...

Fusion

UpsampleA2

...

Encoder1

Decoder1

Encoder2

Decoder2

weights

Diff

Softmax

Differentiator

UpsampleA1

Fusion

UpsampleA2

UpsampleB

FC or

ASPP

FC or

ASPP

Figure 2: Overview of the encoder-decoder structure, which is composed of three parts. (1) Encoder: takes an image as

input and produces multiple feature maps of different resolution. (2) Decoder: gathers feature maps from the encoder and

recovers spatial information. There are two subbranches in the decoder: UpsampleA1 followed by Fusion & UpsampleA2

and UpsampleB. The green and purple double-line arrows traverse one subbranch and the dashed green arrows traverse the

other. One of the two subbranches should be chosen and must be consistent in the two decoder modules. (3) Differentiator:

take the absolute difference of the two recovered feature maps from last part, feed it to a softmax layer, and produce a change

mask.

Unlike CNNs used in object recognition (Simo-

nyan and Zisserman, 2014; He et al., 2016) which

discard all their spatial information and produce high-

level semantic class labels, CNNs for change de-

tection need to predict semantic changes for each

pixel. Thus, the spatial information has to be maintai-

ned, and it’s same as in semantic segmentation. The-

refore, the decoder part is necessary for change de-

tection model.

3.2 Decoder

While encoders are used to produce high-level and

low-level features, decoders gather them to recover

the spatial information and produce the change mask.

In our work, we employ FC (Fully Convolutional

layer) from Fully Convolutional Network(Long et al.,

2015) and ASPP (Atrous Spatial Pyramid Pooling)

from Deeplabv3 (Chen et al., 2017; Chen et al., 2018)

as the main components of our decoders.

FC merely reduces the channel of the input fea-

ture maps while ASPP uses 6-dilated, 12-dilated and

18-dilated convolutional kernels along with a global

pooling layer to extract multi-scale features. As both

of the two modules take the last layers from the en-

coders as input, detailed spatial information such as

boundaries is lost. Therefore, (Long et al., 2015) ups-

amples the feature maps at stride 32, and adds fea-

ture maps at stride 8 from the 3rd pooling layers and

feature maps at stride 16 to them to compensate for

lost spatial information. Similarly, (Chen et al., 2018)

Table 1: Detailed settings for differnet decoders. *A1 refers

to UpsampleA1.

Decoder target size of A1* Fusion

FC stride 16 addition

ASPP stride 4 concatenating

concatenate stride 4 feature maps to upsampled stride

32 feature maps produced by ASPP module.

Following their implementations, we deploy FC

and ASPP as exactly as they do. As Fig. 2 shows,

within the decoder module, two subbranches are pro-

vided. One subbranch fuses feature maps from lo-

wer levels and the other doesn’t. Any one of these

two subbranches can be chosen. Fusion model fuses

low-level features to compensate for detailed boun-

dary spatial information.

Different settings are detailed in Table 1 when FC

or ASPP is deployed. These settings are exactly the

same as the original authors use.

3.3 Differentiator

The siamese encoder-decoder will produce pixel-level

semantic feature maps for the input image pairs. This

provides a basic input with much less interference for

the differentiator. At the uppermost layers of the two

parallel encoder-decoder branches, we take the abso-

lute difference of the two feature maps. This is the dif-

ferencing process, and the formulation is as follows:

i, j,k

= σ(|A

i, j,k

− B

i, j,k

|) (1)

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

528

where L is the feature map produced by the diffe-

rentiator, A and B are feature maps produced by the

two decoders, and σ(·) refers to the softmax layer. k

takes two values here: 0 and 1. Therefore, L

i, j,0

de-

notes the probability of pixel (i, j) being unchanged

and L

i, j,1

denotes the probability of that pixel being

changed.

Then we use the labeled change mask and the

obtained prediction L to train the network in an end-

to-end way with the cross-entropy loss.

In our model, the encoder and decoder compo-

nents are completely decoupled, and hence they can

be replaced by other CNN architectures and combi-

ned to form new networks. In our experiments, we

will demonstrate this point in detail.

4 EXPERIMENTS

4.1 DataSet

We conduct experiments on two datasets: TSUNAMI

and GSV. Both datasets consist of 100 pairs of 224

× 1024 street scene images. The TSUNAMI dataset

contains images of tsunami-damaged areas of Japan,

which are captured by a running vehicle on the street.

The GSV dataset contains Google Street View ima-

ges.

Image pairs in both datasets are coregistered be-

forehand by (Sakurada and Okatani, 2015). For each

image pair, a binary image is provided as the ground

truth, which indicates whether a change occurred for

each corresponding pixel pair. Changes in these two

datasets are deﬁned as changes occurred on the sur-

face of objects (the surface of buildings) and struc-

tural changes (appearing/disappearing objects). The-

refore, grounds, skies, clouds, and illuminations are

not regarded as changes. Further information about

both of the two datasets can be found in (Sakurada

and Okatani, 2015).

An example of changes such as the buildings and

cars are shown in Fig. 1. Both the datasets con-

tain many irrelevant visual differences between image

pairs which meet our demands to train a robust model.

4.2 Conﬁguration

Network structures, training policy, and parameter

settings are detailed in the following paragraphs.

Fine-Tune. VGG16 and Resnet-101, the encoders

in our model, are pretrained on ImageNet. FC, FC-

F, ASPP, and ASPP-F stand for Fully Convolutional

layer, Fully Convolutional layer with Fusion, Atrous

Spatial Pyramid Pooling and Atrous Spatial Pyramid

Pooling with Fusion respectively. Decoders are trai-

ned from scratch except for FC-F, which is ﬁne-tuned

from a trained FC, as FC-F trained from scratch is

slow to converge and hard to outperform its FC coun-

terpart.

Data Augmentation. Due to the insufﬁciency of data,

data augmentation is necessary to train a decent mo-

del. We augment our data in the following three ways.

First, images are left-right ﬂipped randomly with a

probability of 0.5. Second, images are rescaled rand-

omly from 0.5 to 2.0 times the original size. Third,

crop an 800 × 174 patch from the image produced in

the second phase. Before the third phase, an image

might be smaller than 800 × 174 if shrunk too much

in the second phase. In this case, some pixels should

be padded to the right and bottom of the image, in

order to guarantee it’s not smaller than 800 × 174.

Afterward, these pixels will be ignored when the loss

is calculated as they are useless for training.

Optimization. Our models are implemented using

Tensorﬂow and trained on a single NVIDIA TITAN

Xp. We use standard stochastic gradient descent with

batch size of 8, momentum of 0.9 and weight decay of

0.0005 to train our models via back propagation. Each

model is trained for 600 epochs except FC-F is ﬁne-

tuned on a pretrained FC for 1000 epochs (Data aug-

mentation substantially reduces the risk of overﬁtting

if trained for too many epochs). Initial learning rates

for VGG16 and Resnet-101 are set to 0.001 and 0.007

respectively, with the exception of VGG16+FC-F set-

ting to 0.0002. For each model, the learning rate is

controlled by the polynomial learning rate policy:

= α ∗ (1 −

)

(2)

where α

is the learning rate at iteration k, α is the

initial learning rate, k is the current iteration, m is the

number of iterations to ﬁnish the training process and

p is power mentioned above.

In our experiments, we train and evaluate our mo-

dels via 5-fold cross-validation, i.e., 80 image pairs

for training and 20 images for validation, the same

as conﬁgured in (Sakurada and Okatani, 2015). We

report two common metrics from semantic segmen-

tation and machine learning: mean Intersection over

Union (mean IoU) and F1 score. And It can be easily

proved that F1/2 ≤ mean IoU ≤ F1.

4.3 Results and Discussion

Table 2 lists the performance of our models composed

of different encoders and decoders. On TSUNAMI

dataset, our best model VGG16+FC outperforms (Sa-

kurada and Okatani, 2015)’s and (Alcantarilla et al.,

Street-view Change Detection via Siamese Encoder-decoder Structured Convolutional Neural Networks

529

Table 2: Experiment results. *”no aug” means no data augmentation processes are carried out. The sufﬁx *-F means low-level

feature maps are fused. *”Dila” means dilated convolutional layers are used.

Model TSUNAMI mean IoU TSUNAMI F1 score GSV mean IoU GSV F1 score

(Sakurada and Okatani, 2015) - 0.723 - 0.639

(Alcantarilla et al., 2018) - 0.774 - 0.614

VGG16+FC (no aug*) 0.617 0.751 - -

VGG16+FC 0.707 0.819 0.526 0.671

VGG16+FC-F* 0.662 0.788 0.466 0.613

VGG16+ASPP 0.579 0.718 0.409 0.516

VGG16+ASPP-F 0.618 0.752 0.512 0.662

Resnet-101+FC 0.597 0.740 0.531 0.675

Resnet-101+FC-F 0.583 0.722 0.509 0.654

Resnet-101+ASPP 0.477 0.635 0.496 0.642

Resnet-101+ASPP-F 0.569 0.719 0.520 0.661

Resnet-101+Dila*+ASPP 0.634 0.767 0.539 0.681

Resnet-101+Dila+ASPP-F 0.614 0.753 0.545 0.697

2018)’s by 0.096 and 0.045 respectively. On GSV da-

taset, our best model Resnet-101+Dila+ASPP-F out-

performs (Sakurada and Okatani, 2015)’s and (Alcan-

tarilla et al., 2018)’s by 0.058 and 0.083 respectively.

Our best results on TSUNAMI and GSV

are achieved by VGG16+FC and Resnet-

101+Dila+ASPP-F respectively. Results of GSV are

much worse than those of TSUNAMI. We attribute

it to the complexity of GSV. First, there are much

more small objects in GSV than TSUNAMI, such

as people, cars and trees. These small objects are

hard to be identiﬁed by models that lose too much

spatial information such as VGG16+FC. An example

of these disappearing small objects is illustrated in

Fig. 3. Note that cars in the middle of the scene and

branches of trees have totally gone in the prediction

of VGG+FC. Second, boundaries in GSV are sharper.

How to reﬁne boundary is a hard problem in semantic

segmentation, and more smooth boundaries in the

predictions of our models cause the F1 score to be

lower than those in TSUNAMI.

Table 2 also shows the detailed performance of

models constructed with different encoders and de-

coders.

FC and ASPP. ASPP outperforms FC only when di-

lated convolutional layers are present. ASPP captu-

res multi-scale features using large dilated convoluti-

onal kernels. Thus, ASPP makes full of its advantages

when the input feature maps are large. However, the

image heights in our experiments are 174 and 224 in

training and inference process respectively and will be

reduced to 11 and 14. In this case, the size of feature

maps is even smaller than the size of abovementioned

dilated convolutional kernels, and leads to bad perfor-

mance of ASPP in the absence of dilated kernels.

Fusion and Dilation. Fusion improves the perfor-

mance of ASPP largely when dilated kernels are not

used. When dilated kernels are used, the advantages

of Fusion are not so obvious. Fusion introduces spa-

tial information from low-level features and dilated

kernels preserve low-level information. They perform

the same function to some extent. So it explains why

Fusion helps less when dilated kernels are present.

Another notable thing is that data augmentation

is necessary for the robustness of our models, espe-

cially when the dataset is small. As Table 2 shows,

VGG16+FC with augmentation obtains a much better

result than VGG16+FC without augmentation. Fig.

4 partly explains the reason: model training without

augmentation suffers from overﬁtting.

Basically, Resnet-101 combined with ASPP per-

forms much better than VGG16 combined with FC

in semantic segmentation on several datasets in al-

most all aspects. However, VGG16+FC soundly beats

Resnet-101+Dila+ASPP-F on TSUNAMI as shown

in Table 2, which is very counterintuitive. As we trac-

ked the training and validation process, we found it

suffer from overﬁtting. Resnet-101+Dila+ASPP-F is

a complex model and good at recovering ﬁne bounda-

ries. On the one hand, the number of images in TSU-

NAMI is too small to properly ﬁne-tune such a com-

plex model. On the other hand, there are fewer objects

in TSUNAMI than GSV and most objects are large in

size. Although VGG16+FC cannot recover ﬁne boun-

daries of objects, this doesn’t affect the accuracy as

severely as that in GSV.

5 CONCLUSIONS

In this paper, we have proposed a novel approach,

called SEDS-CNN, for street-view change detection.

The SEDS-CNN model is able to handle the irre-

levant visual differences in change detection by in-

troducing the encoder-decoder parts. VGG+FC gi-

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

530

(a) image1

(b) image2

(d) VGG16+FC

(e) VGG16+FC-F

(f) Resnet-101+Dila+ASPP-F

Figure 3: Illustration of our models on one TSUNAMI sample image pair and one GSV sample image pair. The left column

comes from TSUNAMI and the right column comes from GSV.

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

0.4 0.8 1.2 1.6 2.1 2.5 2.9 3.3 3.7 4.1 4.5 4.9 5.3 5.7 6.0

mean IoU

iteration (k)

train val

Figure 4: Training (80 pairs) and validation (20 pairs) mean

IoU over iterations for VGG16+FC with no data augmenta-

tion. It’s suffering from overﬁtting.

ves a decent result when the scenes are simple, ob-

jects are large and datasets are small, while Resnet-

101+Dila+ASPP-F performs better when the opposite

is true. Experiments show that techniques in semantic

segmentaion beneﬁt change detection.

The encoder and decoder parts are decoupled. It is

ﬂexible to choose various CNN architectures as enco-

ders and decoders. Moreover, it is convenient to train

the SEDS-CNN model in an end-to-end way. Ex-

periments on TSUNAMI and GSV datasets demon-

strate that the proposed SEDS-CNN model outper-

forms previous methods by a large margin.

ACKNOWLEDGEMENTS

This work was supported in part by the Natural

Science Foundation of China under Grants U1435220

and 61503365, Youth Innovation Foundation of the

4th China High Resolution Earth Observation Confe-

rence under Grant GFZX04061502, and STS project

of Fujian Province and Chinese Academy of Sciences

under Grant 2018T3009.

Street-view Change Detection via Siamese Encoder-decoder Structured Convolutional Neural Networks

531

REFERENCES

Alcantarilla, P. F., Stent, S., Ros, G., Arroyo, R., and

Gherardi, R. (2018). Street-view change detection

with deconvolutional networks. Autonomous Robots,

42(7):1301–1322.

Bruzzone, L. and Prieto, D. F. (2000). Automatic analy-

sis of the difference image for unsupervised change

detection. IEEE Transactions on Geoscience and Re-

mote sensing, 38(3):1171–1182.

Chen, L.-C., Papandreou, G., Schroff, F., and Adam,

H. (2017). Rethinking atrous convolution for

semantic image segmentation. arXiv preprint

arXiv:1706.05587.

Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., and

Adam, H. (2018). Encoder-decoder with atrous se-

parable convolution for semantic image segmentation.

arXiv preprint arXiv:1802.02611.

Gong, M., Zhao, J., Liu, J., Miao, Q., and Jiao, L. (2016).

Change detection in synthetic aperture radar images

based on deep neural networks. IEEE transactions

on neural networks and learning systems, 27(1):125–

138.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resi-

dual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Li, Y., Martinis, S., Plank, S., and Ludwig, R. (2018). An

automatic change detection approach for rapid ﬂood

mapping in sentinel-1 sar data. International Jour-

nal of Applied Earth Observation and Geoinforma-

tion, 73:123–135.

Lin, T.-Y., Doll

ar, P., Girshick, R. B., He, K., Hariharan, B.,

and Belongie, S. J. (2017). Feature pyramid networks

for object detection. In CVPR, volume 1, page 4.

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu,

C.-Y., and Berg, A. C. (2016). Ssd: Single shot mul-

tibox detector. In European conference on computer

vision, pages 21–37. Springer.

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-

volutional networks for semantic segmentation. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 3431–3440.

Radke, R. J., Andra, S., Al-Kofahi, O., and Roysam, B.

(2005). Image change detection algorithms: a syste-

matic survey. IEEE transactions on image processing,

14(3):294–307.

Redmon, J. and Farhadi, A. (2017). Yolo9000: better, faster,

stronger. arXiv preprint.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Sakurada, K. and Okatani, T. (2015). Change detection

from a street image pair using cnn features and su-

perpixel segmentation. In BMVC, pages 61–1.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Singh, A. (1989). Review article digital change detection

techniques using remotely-sensed data. International

journal of remote sensing, 10(6):989–1003.

Tewkesbury, A. P., Comber, A. J., Tate, N. J., Lamb, A., and

Fisher, P. F. (2015). A critical synthesis of remotely

sensed optical image change detection techniques. Re-

mote Sensing of Environment, 160:1–14.

Yu, F. and Koltun, V. (2015). Multi-scale context ag-

gregation by dilated convolutions. arXiv preprint

arXiv:1511.07122.

Zagoruyko, S. and Komodakis, N. (2015). Learning to com-

pare image patches via convolutional neural networks.

In Proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition, pages 4353–4361.

Zhao, H., Shi, J., Qi, X., Wang, X., and Jia, J. (2017). Py-

ramid scene parsing network. In IEEE Conf. on Com-

puter Vision and Pattern Recognition (CVPR), pages

2881–2890.

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

532