=Paper=
{{Paper
|id=Vol-2259/aics_6
|storemode=property
|title=ADNet: A Deep Network for Detecting Adverts
|pdfUrl=https://ceur-ws.org/Vol-2259/aics_6.pdf
|volume=Vol-2259
|authors=Murhaf Hossari,Soumyabrata Dev,Matthew Nicholson,Killian McCabe,Atul Nautiyal,Clare Conran,Jian Tang,Wei Xu,Francois Pitie
|dblpUrl=https://dblp.org/rec/conf/aics/HossariDNMNCTXP18
}}
==ADNet: A Deep Network for Detecting Adverts==
https://ceur-ws.org/Vol-2259/aics_6.pdf
ADNet: A Deep Network for Detecting Adverts
Murhaf Hossari?1 , Soumyabrata Dev?1 , Matthew Nicholson1 , Killian McCabe1 ,
Atul Nautiyal1 , Clare Conran1 , Jian Tang3 , Wei Xu3 , and François Pitié1,2
1
The ADAPT SFI Research Centre, Trinity College Dublin
2
Department of Electronic & Electrical Engineering, Trinity College Dublin
3
Huawei Ireland Research Center, Dublin
Abstract. Online video advertising gives content providers the ability
to deliver compelling content, reach a growing audience, and generate ad-
ditional revenue from online media. Recently, advertising strategies are
designed to look for original advert(s) in a video frame, and replacing
them with new adverts. These strategies, popularly known as product
placement or embedded marketing, greatly help the marketing agencies
to reach out to a wider audience. However, in the existing literature,
such detection of candidate frames in a video sequence for the purpose
of advert integration, is done manually. In this paper, we propose a deep-
learning architecture called ADNet, that automatically detects the pres-
ence of advertisements in video frames. Our approach is the first of its
kind that automatically detects the presence of adverts in a video frame,
and achieves state-of-the-art results on a public dataset.
Keywords: deep learning · advertisements · ADNet · CNN · product
placement.
1 Introduction
With the ubiquity of multimedia videos, there has been a massive interest from
advertising and marketing agencies to provide targeted advertisements for cus-
tomers. Such targeted adverts are useful, both from the perspectives of market-
ing agents and end-users or consumers. The rise of television and online videos
has provided several avenues in transmitting brand information to the audience.
Several studies [14, 8] have shown that there is a steady growth in the gener-
ated revenues from internet advertising over the last decade, and the growth is
forecast to continually increase in the coming years.
Traditionally, in online videos, advertisement videos are inserted within the
online video content. The advertisement videos are played independently, either
prior the online content (pre-roll); or during the online content (mid-roll); or
after the online content is viewed by the user (post-roll). The option of fast-
forwarding the advertisement video is disabled. Such adverts hugely interrupt
the seamless viewing experiences of the viewers [10]. Therefore, nowadays, adver-
tisement placements are integrated directly into the scene. They are popularly
?
Authors contributed equally and arranged alphabetically.
2 M. Hossari, S. Dev et al.
known as embedding marketing, or product placements, where specific brands or
products are deliberately inserted into the scene [9]. Such placements are mostly
manually done via digital editing technologies in the post-processing phase. This
involves manually going through all possible frames of the video, and identifying
candidate video frames for advert integration. The existing adverts are replaced
with new adverts, in order to reach out to specific demographic population or
markets.
In this paper, we automatically identify the frames in a video sequence that
have an existing advert. Such an automatic process will greatly help during the
post-processing stage, as the video editor no longer needs to parse the video
frames manually. Our innovative AI-powered system allows the detection of ad-
verts in video sequences. Such system will greatly expedite the process of post-
processing, and save a huge amount of man-hours.
In this work, we focus our attention on detecting billboards in outdoor scene
videos. We propose a deep learning architecture called ADNet, that automat-
ically detects billboard adverts 4 in a video sequence. Our proposed model is
inspired from the state-of-the-art architectures, in creating a bespoke network
that can automatically detect existing billboards in a video. The rest of the pa-
per is organized as follows: Section 2 discusses the related work. In Section 3, we
describe the ADNet architecture and its layer configuration in a detailed man-
ner. We train the ADNet model using a composite dataset consisting of positive-
(images containing a billboard) and negative- (images that do not contain any
billboard) examples. Section 4 describes the details of the dataset, and the bench-
marking results. Finally, we conclude the paper in Section 5, and mention our
future work.
2 Related Work
In the recent years, deep convolutional neural networks have proven to be very
effective in the areas of image- and video- processing. More specifically, they
have been widely used to tackle tasks related to image classification and object
detection. The convolutional neural networks, popularly known as ConvNets are
a type of feed-forward neural nets, that produce state-of-the-art results in the ar-
eas of visual recognition. There are several reasons for these massive successes in
this area. Firstly, the public repository of large-scale image- and video- datasets
have helped the machine learning researchers to train the deep neural networks
on millions of images [5]. The ImageNet Large-Scale Visual Recognition Chal-
lenge (ILSVRC) is a good example [15], that led to the development of several
novel neural networks. Secondly, the rise of high computing machines viz. graph-
ics processing unit (GPU) and clusters [4] helped the researchers to train deep
networks on very large datasets.
The existing works in advert detection primarily revolve around the detection
of advertisement clips in a video sequence. Recently, Hussain et al. in [7] propose
4
We will use the terms advert and billboard interchangeably throughout the paper.
ADNet: A Deep Network for Detecting Adverts 3
a novel solution for automatically understanding the advertisement content, and
analyzing the sentiments around them. Feng et al. worked on the detection of
logos in television commercials [6], using a combination of audio and video fea-
tures. Acoustic match profiles are also exploited in [3] to accurately locate the
adverts in the video sequence. These works on advert detection do not consider
identifying a particular object in the video scene, and subsequently integrating
a new advert into the same scene. However, commercial companies viz. Mirriad
uses patented technology to insert new objects into a video scene.
In the area of sport analytics, a few works were done to detect the billboards
in the soccer fields. Watve et al. attempted to localize the position of on-field
billboards using template matching techniques [18]. Cai et al. in [2] used Hough
transform for advertising detection in sport TV. Aldershoff et al. used photo-
metric invariant features for billboard track in soccer video scenes [1]. These
techniques mainly used techniques from photogrammetry and traditional signal
processing – neural networks were not fully exploited. Owing to the recent de-
velopment in the area of artificial intelligence, we propose a AI-driven advert
detection neural network in this paper.
3 ADNet architecture
Our proposed advert-detection model, ADNet (cf. Fig 1) is inspired from the
VGG19 architecture. The VGG19 model uses very small convolution filters (3 ×
3), with depth upto 19 weight layers. The VGG19 has 6 different configurations,
depending on the number of weight layers. These configurations are denoted by
A, A-LRN, B, C, D, and E. More details on this can be found in [16].
Fig. 1: Architecture of proposed ADNet.
The ADNet uses the pre-trained weights of the VGG network, trained on
the ImageNet dataset. We toggle off the training for the first 5 layers of the pre-
trained VGG network. We also remove the last 3 weight layers of the original
VGG19 model, and add a stack of Fully-Connected (FC) layers. The first FC-
layer has 1024 channels with relu activation function. In order to prevent the
4 M. Hossari, S. Dev et al.
ADNet configuration
A A-LRN B C D E
11 weight 11 weight 13 weight 16 weight 16 weight 19 weight
layers layers layers layers layers layers
input (224 X 224 RGB image)
conv3-64 conv3-64 conv3-64 conv3-64 conv3-64 conv3-64
LRN conv3-64 conv3-64 conv3-64 conv3-64
maxpool
conv3-128 conv3-128 conv3-128 conv3-128 conv3-128 conv3-128
conv3-128 conv3-128 conv3-128 conv3-128
maxpool
conv3-256 conv3-256 conv3-256 conv3-256 conv3-256 conv3-256
conv3-256 conv3-256 conv3-256 conv3-256 conv3-256 conv3-256
conv1-256 conv3-256 conv3-256
conv3-256
maxpool
conv3-512 conv3-512 conv3-512 conv3-512 conv3-512 conv3-512
conv3-512 conv3-512 conv3-512 conv3-512 conv3-512 conv3-512
conv1-512 conv3-512 conv3-512
conv3-512
maxpool
conv3-512 conv3-512 conv3-512 conv3-512 conv3-512 conv3-512
conv3-512 conv3-512 conv3-512 conv3-512 conv3-512 conv3-512
conv1-512 conv3-512 conv3-512
conv3-512
FC-1024
Dropout (0.5)
FC-1024
FC-2
Table 1: Layer configuration of ADNet model. The ADNet model is inspired
from the VGG19 model – the distinct additional layers in ADNet are indicated
in blue color. The convolutional layers are indicated as conv{receptive field size}–
{number of channels}. The fully connected layers are indicated as FC-{number
of channels}. The dropout layer is indicated as Dropout ({rate of dropout}).
ADNet from over-fitting, we add a dropout layer with a rate of 0.5. The second
FC-layer has also 1024 channels with relu activation function. The third and
final FC-layer has 2 channels, with softmax activation function. The two values
obtained at the end of ADNet, provide the probabilities of belonging to billboard
or no-billboard classes.
We describe the layer configuration of the ADNet architecture in Table 1 in
a detailed manner. Similar to VGG model in [16], we experiment with various
configurations of ADNet viz. A, A-LRN, B, C, D, E. As expected, we get the
best performance for advert detection with the configuration E, consisting of 19
weight layers. This makes sense, as the neural network becomes more discrimi-
native with deeper layers, and provides better classification performance. In the
ADNet: A Deep Network for Detecting Adverts 5
subsequent discussion, the default configuration of ADNet is configuration E,
with 19 weight layers.
4 Experiments and Results
4.1 Dataset
We train our ADNet model on a composite dataset that consists of both positive-
and negative- examples of billboard detection. The positive examples are those
outdoor-scene images that contain a billboard. The billboard object should cover
a significant proportion of the image. We set the threshold as follows – the
billboard should cover more than 10% of the total image area. The images from
this composite dataset are collected from two sources:
– Mapillary Vistas Dataset [13]
– Microsoft COCO (MS-COCO) Dataset [11]
Our composite dataset has a total of 18945 images. Out of these images,
9141 images contain billboard images; and the remaining 9804 images do not
contain any billboard. Most of the positive examples are images from the Map-
illary Vistas Dataset. We include those images from the dataset, that contain
billboard(s) in them. Any images that contain billboard(s), covering less than
10% of the total image area are ignored. We also exclude images, where the
billboards are partially off-screen. Most of the negative examples (i.e. images
with no billboards) are selected from Microsoft COCO Dataset. We randomly
selected 9395 images from MS-COCO that do not contain any adverts.
We divide the entire 18945 images of the composite dataset into training-
and testing- sets. Table 2 describes the distribution of images into training- and
testing- sets.
Positive Negative
Set Total
Mapillary MS-COCO Mapillary MS-COCO
Training 6380 - 303 6617 13300
Testing 2761 - 106 2778 5645
Total 9141 - 409 9395 18945
Table 2: Distribution of training- and testing- sets, with respect to the two
datasets.
4.2 Subjective Evaluation
In this section, we provide a subjective evaluation of the advert detection using
ADNet model. Our large-scale composite dataset contains a balanced number of
positive- and negative- billboard examples. Figure 2 shows a few visual results.
6 M. Hossari, S. Dev et al.
(a) Positive examples of adverts correctly identified by ADNet.
(b) Negative examples of adverts correctly ignored by ADNet.
(c) Examples of misclassification errors.
Fig. 2: A few visual illustrations of ADNet performance.
ADNet: A Deep Network for Detecting Adverts 7
Figure 2(a) shows a few sample results that are correctly classified by the AD-
Net model. All of them are outdoor scenes that contain billboard(s) at prominent
location of the scene. Figure 2(b) are a few images that do not contain adverts,
and are correctly ignored by ADNet. Finally, we show a few misclassification
errors of ADNet in Fig 2(c). Most of these images contain objects that are sim-
ilar in shape to a regular four-sided billboard. We observe that the back of the
truck, the solar cell stand and the road signage has similar shape and appearance
to a regular billboard. Therefore, the ADNet model predicts that these images
contain regular billboard in them.
4.3 Objective Evaluation
In order to provide an objective evaluation of our model, we benchmark the
performance of ADNet model with the Inception model. We used Inception-v3
and used the pre-trained weights that were obtained by training the model for
ImageNet Large Visual Recognition Challenge 2012. This recognition challenge
involved classifying the images into 1000 predefined classes. The Inception-v3
network consists of several blocks of ‘inception’ layers, which are a combination of
multiple convolutional and pooling layers [17]. We use the keras-implementation
of Inception-v3 that uses 311 layers. We froze the training for the first 172 layers,
and toggled the training ‘on’ for the remaining layers. We also added an average
pooling layer, and a fully connected layer with 1024 channels (FC-1024). Finally,
we added a fully connected layer with 2 channels (FC-2) with softmax activation.
Similar to ADNet, this Inception-v3 model provides us two values that indicate
of probability of billboard or no-billboard classes.
We trained our ADNet model on the composite dataset of 18945 images,
for 50 epochs with stochastic gradient descent optimizer. We set the learning
rate as 0.0001, with a batch size of 16, and using categorical cross-entropy loss
function. We used the same configuration file to train the Inception-v3 model on
the composite image dataset.
In this paper, we report the classification accuracy of ADNet. Suppose T P ,
T N , F P and F N denote the true positive, true negative, false positive and false
negative samples of our binary classification of billboard detection. We define
the classification accuracy for this task as:
TP + TN
Classification Accuracy = . (1)
TP + TN + FP + FN
Table 3 describes the performance evaluation of our proposed model, with
the state-of-the-art Inception model.
Approach Classification Accuracy
Inception-v3 0.56
Proposed model ADNet 0.94
Table 3: Performance evaluation of our proposed model in detecting adverts.
8 M. Hossari, S. Dev et al.
The ADNet model has a competitive classification accuracy of 0.94, as com-
pared to the benchmarking approaches. Moreover, it has a competitive processing
speed on video frames. In a Linux computer with a GPU configuration of Nvidia
GTX1080Xtreme D5X 8GB, the computational time of ADNet during testing
stage is 57ms for every frame in the video. Such competitive results on real-life
video frames, opens up the possibility of deploying such detection model in a
real-time advert creation system [12].
5 Conclusion
In this paper, we have proposed a deep neural network called ADNet, that can
automatically parse the frames of a video, and identify the frames that contain
an advert. This is useful for the advertisement agencies and marketers, to provide
targeted advertisements for specific markets and demographics. The ADNet can
help the user, in identifying the key frames in the video where new adverts can be
integrated. Our approach is the first of its kind that uses deep neural networks
for detecting adverts in video frames. In our future work, we plan to extend
ADNet by relaxing the criterion of outdoor scene images. We will explore its
possibility of detecting adverts, for indoor-scene images viz. movie and reality
television shows.
Acknowledgement
The ADAPT Centre for Digital Content Technology is funded under the SFI
Research Centres Programme (Grant 13/RC/2106) and is co-funded under the
European Regional Development Fund.
References
1. Aldershoff, F., Gevers, T.: Visual tracking and localization of billboards in streamed
soccer matches. In: Storage and Retrieval Methods and Applications for Multime-
dia 2004. vol. 5307, pp. 408–417. International Society for Optics and Photonics
(2003)
2. Cai, G., Chen, L., Li, J.: Billboard advertising detection in sport tv. In: Signal Pro-
cessing and Its Applications, 2003. Proceedings. Seventh International Symposium
on. vol. 1, pp. 537–540. IEEE (2003)
3. Covell, M., Baluja, S., Fink, M.: Advertisement detection and replacement using
acoustic and visual repetition. In: Multimedia Signal Processing, 2006 IEEE 8th
workshop on. pp. 461–466. IEEE (2006)
4. Dean, J., Corrado, G., Monga, R., Chen, K., Devin, M., Mao, M., Senior, A.,
Tucker, P., Yang, K., Le, Q.V., et al.: Large scale distributed deep networks. In:
Advances in neural information processing systems. pp. 1223–1231 (2012)
5. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-
Scale Hierarchical Image Database. In: CVPR09 (2009)
6. Feng, Z., Neumann, J.: Real time commercial detection in videos (2013)
ADNet: A Deep Network for Detecting Adverts 9
7. Hussain, Z., Zhang, M., Zhang, X., Ye, K., Thomas, C., Agha, Z., Ong, N., Ko-
vashka, A.: Automatic understanding of image and video advertisements. In: 2017
IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1100–
1110. IEEE (2017)
8. IAB, P.: Iab internet advertising revenue report 2011 full-year results. Tech. rep.,
Market research report, Interactive Advertising Bureau (IAB) and Pricewater-
houseCoopers (PwC). http://www. iab. net/media/file/IAB Internet Advertising
Revenue Report FY 2011. pdf (2012)
9. Karniouchina, E.V., Uslay, C., Erenburg, G.: Do marketing media have life cycles?
the case of product placement in movies. Journal of Marketing 75(3), 27–48 (2011)
10. Li, H., Lo, H.Y.: Do you recognize its brand? the effectiveness of online in-stream
video advertisements. Journal of Advertising 44(3), 208–218 (2015)
11. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P.,
Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference
on computer vision. pp. 740–755. Springer (2014)
12. Nautiyal, A., McCabe, K., Hossari, M., Dev, S., Nicholson, M., Conran, C., McK-
ibben, D., Tang, J., Wei, X., Pitié, F.: An advert creation system for next-gen
publicity. In: Proc. European Conference on Machine Learning and Principles and
Practice of Knowledge Discovery in Databases (ECML-PKDD) (2018)
13. Neuhold, G., Ollmann, T., Bulò, S.R., Kontschieder, P.: The mapillary vistas
dataset for semantic understanding of street scenes. In: ICCV. pp. 5000–5009
(2017)
14. PricewaterhouseCoopers, L.: Iab internet advertising revenue report-2012 full year
results. Bericht, Interactive Advertising Bureau (2013)
15. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z.,
Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large
Scale Visual Recognition Challenge. International Journal of Computer Vision
(IJCV) 115(3), 211–252 (2015). https://doi.org/10.1007/s11263-015-0816-y
16. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. CoRR abs/1409.1556 (2014)
17. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the incep-
tion architecture for computer vision. In: Proceedings of the IEEE conference on
computer vision and pattern recognition. pp. 2818–2826 (2016)
18. Watve, A., Sural, S.: Soccer video processing for the detection of advertisement
billboards. Pattern Recognition Letters 29(7), 994–1006 (2008)