Professional Documents
Culture Documents
PDF Advances in Computer Vision Proceedings of The 2019 Computer Vision Conference CVC Volume 1 Kohei Arai Ebook Full Chapter
PDF Advances in Computer Vision Proceedings of The 2019 Computer Vision Conference CVC Volume 1 Kohei Arai Ebook Full Chapter
PDF Advances in Computer Vision Proceedings of The 2019 Computer Vision Conference CVC Volume 1 Kohei Arai Ebook Full Chapter
https://textbookfull.com/product/proceedings-of-the-future-
technologies-conference-ftc-2018-volume-1-kohei-arai/
https://textbookfull.com/product/proceedings-of-the-future-
technologies-conference-ftc-2020-volume-1-kohei-arai/
https://textbookfull.com/product/intelligent-computing-
proceedings-of-the-2020-computing-conference-volume-1-kohei-arai/
https://textbookfull.com/product/applied-computer-vision-and-
image-processing-proceedings-of-iccet-2020-volume-1-brijesh-iyer/
Intelligent Computing: Proceedings of the 2018
Computing Conference, Volume 2 Kohei Arai
https://textbookfull.com/product/intelligent-computing-
proceedings-of-the-2018-computing-conference-volume-2-kohei-arai/
https://textbookfull.com/product/proceedings-of-the-future-
technologies-conference-ftc-2018-volume-2-kohei-arai/
https://textbookfull.com/product/intelligent-computing-
proceedings-of-the-2020-computing-conference-volume-3-kohei-arai/
https://textbookfull.com/product/intelligent-computing-
proceedings-of-the-2020-computing-conference-volume-2-kohei-arai/
https://textbookfull.com/product/computer-vision-systems-12th-
international-conference-icvs-2019-thessaloniki-greece-
september-23-25-2019-proceedings-dimitrios-tzovaras/
Advances in Intelligent Systems and Computing 943
Kohei Arai
Supriya Kapoor Editors
Advances in
Computer
Vision
Proceedings of the 2019 Computer
Vision Conference (CVC), Volume 1
Advances in Intelligent Systems and Computing
Volume 943
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, Electronic Engineering, University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen, Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
Editors
Advances in
Computer Vision
Proceedings of the 2019 Computer Vision
Conference (CVC), Volume 1
123
Editors
Kohei Arai Supriya Kapoor
Saga University The Science and Information
Saga, Saga, Japan (SAI) Organization
Bradford, West Yorkshire, UK
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
It gives us the great pleasure to welcome all the participants of the Computer Vision
Conference (CVC) 2019, organized by The Science and Information
(SAI) Organization, based in the UK. CVC 2019 offers a place for participants to
present and to discuss their innovative recent and ongoing research and their
applications. The prestigious conference was held on 25–26 April 2019 in
Las Vegas, Nevada, USA.
Computer vision is a field of computer science that works on enabling the
computers to identify, see and process information in a similar way that humans do
and provide an appropriate result. Nowadays, computer vision is developing at a
fast pace and has gained enormous attention.
The volume and quality of the technical material submitted to the conference
confirm the rapid expansion of computer vision and CVC’s status as its flagship
conference. We believe the research presented at CVC 2019 will contribute to
strengthen the great success of computer vision technologies in industrial, enter-
tainment, social and everyday applications. The participants of the conference were
from different regions of the world, with the background of either academia or
industry.
The published proceedings has been divided into two volumes, which covered a
wide range of topics in Machine Vision and Learning, Computer Vision
Applications, Image Processing, Data Science, Artificial Intelligence, Motion and
Tracking, 3D Computer Vision, Deep Learning for Vision, etc. These papers are
selected from 371 submitted papers and have received the instruction and help from
many experts, scholars and participants in proceedings preparation. Here, we would
like to give our sincere thanks to those who have paid great efforts and support
during the publication of the proceeding. After rigorous peer review, 118 papers
were published including 7 poster papers.
Many thanks go to the Keynote Speakers for sharing their knowledge and
expertise with us and to all the authors who have spent the time and effort to
contribute significantly to this conference. We are also indebted to the organizing
committee for their great efforts in ensuring the successful implementation of the
v
vi Preface
conference. In particular, we would like to thank the technical committee for their
constructive and enlightening reviews on the manuscripts in the limited timescale.
We hope that all the participants and the interested readers benefit scientifically
from this book and find it stimulating in the process. See you in next SAI
Conference, with the same amplitude, focus and determination.
Regards,
Kohei Arai
Contents
vii
viii Contents
1 Introduction
The ever increasing modernisation of signal systems and electrification of major
railway lines lead to increasingly complex railway environments. These environ-
ments require advanced management systems which incorporate a detailed rep-
resentation of the network, its assets and surroundings. The aim of such systems
is to facilitate automation of processes and improved decision support, minimis-
ing the requirement for expensive and inefficient on-site activities. Fundamental
requirements are detailed maps and databases of railway assets, such as poles,
signs, wires, switches, cabinets, signalling equipment, as well as the surrounding
environment including trees, buildings and adjacent infrastructure. The detailed
maps may also form the basis for a simulation of the railway as seen from the
train operator’s viewpoint. Such simulations/videos are used in the training of
train operators and support personnel. Ideally, the maps should be constantly
updated to ensure currency of the databases as well as to facilitate detailed
documentation and support of maintenance and construction processes in the
c Springer Nature Switzerland AG 2020
K. Arai and S. Kapoor (Eds.): CVC 2019, AISC 943, pp. 1–15, 2020.
https://doi.org/10.1007/978-3-030-17795-9_1
2 G. Karagiannis et al.
2 Literature Review
with the track bed. The overall average detection achieved by this method is
96.4%. The main drawback of this approach is that it depends on a sophisticated
type of data that needs special equipment to capture and it is more complicated
to process.
Agudo et al. in [1] presented a real-time railway speed limit and warning
signs recognition method on videos. After noise removal, Canny edge detection
is applied and an optimised Hough voting scheme detects the sign region of
interest on the edge image. Based on gradient directions and distances of points
on the edge of a shape, candidate central points of signs occur. Then, recognition
is achieved by applying shape criteria since the signs are either circular, square
or rectangular. The method scored 95.83% overall accuracy on classification.
However, even though the dataset had more than 300,000 video frames, only
382 ground truth signs existed and the authors do not provide any score for the
detection accuracy.
3 Data Analysis
3.1 Dataset
In our case, the dataset consists of 47,912 images acquired in 2013 and show the
railway from Brisbane to Melbourne in Australia. The images were acquired in
three different days in the morning with similar sunlight conditions. The camera
used is a Ladybug3 spherical camera system and the images are panoramic views
of size 5400 × 2700 pixels. The images were processed manually by the produc-
tion team of the company2 for annotations and resulted in 121,528 instances of
railway signs and signals.
2
Second Affiliation.
4 G. Karagiannis et al.
Fig. 1. Instances of signals and signs. First row from left to right (class name in paren-
theses when it is different from the description): speed sign (Sign S), unit sign (Sign
U), speed standard letter (Sign Other), position light main (PL 2W), position light
separately (PL 2W& 1R), signal with two elements (Signal2 F). Second row: diverging
left-right speed sign (Sign LR), diverging left speed sign (Sign L), signal number (Sign
Signal), signal with one element (Signal1 F), signal with three elements (Signal3 F),
signal with four elements (Signal4 F).
Fig. 2. Instances of speed signs at different lighting conditions, viewpoints and scales.
The samples are originally separated into twenty five classes. Each class is in
fact a subclass of two parent classes, signs and signals. Fifteen classes correspond
to signals: three different types of position lights with their back and side view,
signals with one, two, three or four elements (lights), their side and back view
and other type of signals. Also, ten classes correspond to different types of signs:
speed signs, diverging left speed signs, diverging right speed signs, diverging
left-right speed signs, unit signs, signal number signs, other signs, back views of
circular signs, back views of diamond signs and back views of rectangular signs.
From the total amount of samples, 67,839 correspond to signals and 53,689 to
signs. Figure 1 shows some instances of signs and signals. Each of these instances
correspond to a different class. We can see the high similarity among the classes
both for signs and signals. Specifically, diverging left, diverging right, diverging
left-right and regular speed signs are very similar and especially when they are
smaller than the examples shown in Fig. 1. Similarly, even for humans it is often
hard to distinguish between signals with three or four elements when they are
small. All examples shown here are larger than their average size on the dataset
for clarity.
Figure 2 shows examples of regular speed signs with different viewpoint, size
and illumination. These examples illustrate the scale, viewpoint and illumination
variation of the objects but at the same time the high similarity among the
classes.
Railway Signs and Signals 5
Fig. 3. Instances of signals with one element (Signal1). All four examples belong to
the same class even though they have different characteristics. From left to right: some
have long (first and second), no top cover at all (third) or short cover (last). Also, some
have no back cover (first), other have circular (second and third) or semicircular (last)
Fig. 4. Amount of sample instances per class before merging some classes. The imbal-
ance among the classes is too high.
Fig. 5. Amount of sample instances per class. After merging some classes with similar
characteristics we end up with a more balanced dataset.
Fig. 6. Amount of samples according to their size in pixel2 . Most samples (89%) are
smaller than 502 pixels.
A reason behind the high amount of very small objects in our dataset is that
the data was acquired by driving on a single track. However, in many sectors
along the railway there are multiple parallel tracks and the signs and signals
corresponding to these tracks appear in the dataset only in small sizes since
the camera never passed close to them. One way to limit the small object size
problem in our dataset is to split each panoramic image into smaller patches with
high overlap, small enough to achieve a less challenging relative size between
objects and image. Specifically, in our approach, each image is split into 74
patches of size 6002 pixels with 200 pixels overlap on each side. Even at this
level of fragmentation, a 502 object corresponds to 0.69% of a patch size. A
consequence of splitting the images into smaller patches is that the same object
may now appear in more than one patches due to overlap. In fact, while on the
panoramic images there exist 121,528 object instances, on the patches that were
extracted, there exist 203,322 instances. The numbers shown in Fig. 6 correspond
to the objects existing on the patches.
4 Methodology
For the detection of signs and signals, we applied the Faster R-CNN presented
by Ren et al. in [2] using ResNet-101 [21] as feature extractor. We decided to
implement this approach mainly motivated by its high performance on competi-
tive datasets such as Pascal VOC 2007 (85.6% mAP), Pascal VOC 2012 (83.8%
mAP) and COCO (59% mAP). The main drawback of this approach compared
to other acknowledged object detection methods such as YOLO [16]) or SSD is
its high processing time (three to nine times slower depending on the implemen-
tation [17]). However, the sacrifice in time pays off in accuracy, especially in this
dataset since this method performs better on small objects [2].
Here we will present some key points of Faster R-CNN. First, Faster R-
CNN is the descendant of Fast R-CNN [2] which in turn is the descendant of
R-CNN [13]. As their names imply, Fast and Faster R-CNNs are more efficient
implementations of the original concept in [13], R-CNN. The main elements of
Faster R-CNN are: (1) the base network, (2) the anchors, (3) the Region Proposal
Network (RPN) and (4) the Region based Convolutional Neural Network (R-
CNN). The last element is actually Fast R-CNN, so with a slight simplification
we can state that Faster R-CNN = RPN + Fast R-CNN.
The base network is a, usually deep, CNN. This network consists of multi-
ple convolutional layers that perform feature extraction by applying filters at
different levels. A common practice [2,14,16] is to initialize training using a
pre-trained network as a base network. This helps the network to have a more
realistic starting point compared to random initialization. Here we use ResNet
[21]. The second key point of this method is the anchors, a set of predefined
possible bounding boxes at different sizes and aspect ratios. The goal of using
the anchors is to catch the variability of scales and sizes of objects in the images.
Here we used nine anchors consisting of three different sizes (152 , 302 and 602
pixels) and three different aspect ratios 1:1, 1:2 and 2:1.
Next, we use RPN, a network trained to separate the anchors into fore-
ground and background given the Intersection over Union (IoU) ratio between
the anchors and a ground-truth bounding box (foreground if IoU > 0.7 and
background if IoU < 0.1). Thus, only the most relevant anchors for our dataset
are used. It accepts as input the feature map output of the base model and
creates two outputs: a 2 × 9 box-classification layer containing the foreground
and background probability for each of the nine different anchors and a 4 × 9
box-regression layer containing the offset values on x and y axis of the anchor
Railway Signs and Signals 9
Fig. 7. Right: the architecture of Faster R-CNN. Left: the region proposal network
(RPN). Source: Figs. 2 and 3 of [2].
For the evaluation of detection, an overlap criterion between the ground truth
and predicted bounding box is defined. If the Intersection over Union (IoU) of
these two boxes is greater than 0.5, the prediction is considered as True Positive
(TP). Multiple detections of the same ground truth objects are not considered
true positives, each predicted box is either True-Positive or False-Positive (FP).
Predictions with IoU smaller than 0.5 are ignored and count as False Negatives
(FN). Precision is defined as the fraction of the correct detections over the total
10 G. Karagiannis et al.
detections ( T PT+F
P
P ). Recall is the fraction of correct detections over the total
amount of ground truth objects ( T PT+F P
N ) [22]. For the evaluation of classifica-
tion and overall accuracy of our approach, we adopted mean Average Precision
(mAP) [23] as a widely accepted metric [22]. For each class, the predictions
satisfying the overlap criterion are assigned to ground truth objects in descend-
ing order by the confidence output. The precision/recall curve is computed and
the average precision(AP) is the mean value of interpolated precision at eleven
equally spaced levels of recall [22]:
1
AP = pinterp (r) (1)
11
r∈{0,0.1,...,1}
where
pinterp (r) = maxr̃:r̃≥r p(r̃) (2)
Then, the mean of all APs across the classes is the mAP metric. This metric
was used as evaluation method for the Pascal VOC 2007 detection challenge
and from that time is the most common evaluation metric on object detection
challenges [18].
5 Results
The network was trained on a single Titan-X GPU implementation for approx-
imately two days. We trained it for 200k iterations with a learning rate of 0.003
Fig. 8. Percentage of objects detected according to their size in pixels2 . The algorithm
performed very well on detecting objects larger than 400 pixels (more than 79% in
all size groups). On the other hand, it performed poorly on very small objects (less
than 200 pixels) detecting only 24% of them. The performance was two times better
on objects with size 200–400 pixels (57%), which is the most dominant size group with
about 45,000 samples (see Fig. 6).
Railway Signs and Signals 11
and for 60k iterations with learning rate 0.0003. The model achieved an overall
precision of 79.36% on the detection task and a 70.9% mAP.
The precision level of detection is considered high given the challenges
imposed by the dataset as they are presented in Sect. 3. Figure 8 shows the
detection performance of the algorithm with respect to the size of the samples.
We can see that the algorithm fails to detect very small objects. About 76%
of objects smaller than 200 pixels are not detected. A significantly better, but
still low, performance is observed for objects of 200–400 pixels size (57% detec-
tion rate). On the other hand, the algorithm performs uniformly well for objects
Fig. 9. Example of detection. Two successful detections and one false positive. We
can see the illumination variance even in a single image with areas in sunlight and in
shadow. The sign is detected successfully despite its small size and the partial occlusion
from the bush. The entrance of the tunnel was falsely detected as a signal with three
elements.
12 G. Karagiannis et al.
larger than 400 pixels (79% for sizes 400–600 and more than 83% for larger than
600 pixels). These results show that, in terms of object size, there is a threshold
above which the performance of Faster-RCNN is stable and it is unaffected by
the size of the objects. Therefore, if there are enough instances and the objects
are large enough, object size is not a crucial factor. In our case, this threshold
is about 500 pixels.
An interesting metric for our problem would be to measure the amount of
unique physical objects detected. The goal of object detection in computer vision,
usually, is to detect all the objects that appear in an image. In mapping, most
of the times, it is important to detect the locations of the physical objects. If we
have multiple images showing the same physical object from different point of
views and at different scales, in mapping, it would be sufficient to detect it at
Fig. 10. Example of detection. A successful detection of more complex object struc-
tures. Different signals mounted on top of each other did not confuse the algorithm.
Railway Signs and Signals 13
least in one image and not necessarily in all images. We expect that our approach
will perform better in this metric, however, information about unique physical
objects is not available for this dataset.
Figures 9 and 10 show two representative examples of the results. We can see
in these figures the variance in illumination conditions and the small size of the
objects. The speed signs that appear on these images are very small but they
belong to the most common object size interval for this dataset (200–400 pixels).
By looking at these figures, we can also realise the difficulty of the classification
task of this dataset. Even for a human, it is hard to decide on which class the
objects on these images belong to.
Fig. 11. Confusion matrix showing the classification AP of the detected objects. Rows
represent the predictions and columns the ground truth objects. Last row and last
column (red colour) summarise the performance per class. Bottom right cell (blue
colour) is the mAP of the method. See Fig. 2 for class name explanation.
14 G. Karagiannis et al.
6 Conclusions
In this paper, we presented a novel dataset used in real world projects for map-
ping of railway infrastructure. The dataset is very challenging mainly because
the objects are very small relatively to the size of the entire images but also
absolutely in terms of pixels2 . We implemented a state of the art deep learn-
ing method for detection and classification of signs and signals on this dataset
scoring 79.36% on the detection task and 70.9% on the classification task. We
believe that the accuracy of the model will be higher in a similar dataset with
better resolution images. Also, given that the object positions along a railway fol-
low specific regulations, taking into account the spatial relationships among the
objects can improve accuracy. Moreover, it would be interesting to apply other
state of the art methods on this dataset and analyse the strengths and weak-
nesses of each method on the same tasks. In addition, for mapping tasks such
as the current, a more appropriate metric would be to evaluate detections based
on the amount of physical objects and not on objects existing in the images. In
conclusion, our results are promising and suggest further investigation into the
use of deep learning for railway asset detection and mapping.
References
1. Agudo, D., Sánchez, Á., Vélez, J.F., Moreno, A.B.: Real-time railway speed limit
sign recognition from video sequences. In: 2016 International Conference on Sys-
tems, Signals and Image Processing (IWSSIP), pp. 1–4, May 2016
2. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object
detection with region proposal networks. In: Advances in Neural Information Pro-
cessing Systems, vol. 2015, pp. 91–99 (2015)
3. Zheng, D., Wang, Y.: Application of an artificial neural network on railway passen-
ger flow prediction. In: Proceedings of 2011 International Conference on Electronic
Mechanical Engineering and Information Technology, vol. 1, pp. 149–152, August
2011
4. Tsai, T.-H., Lee, C.-K., Wei, C.-H.: Neural network based temporal feature mod-
els for short-term railway passenger demand forecasting. Expert Syst. Appl.
36(2), 3728–3736 (2009). Part 2. http://www.sciencedirect.com/science/article/
pii/S0957417408001516
Railway Signs and Signals 15
5. Karthick, N., Nagarajan, R., Suresh, S., Prabhu, R.: Implementation of railway
track crack detection and protection. Int. J. Eng. Comput. Sci. 6(5), 21476–21481
(2017). http://ijecs.in/index.php/ijecs/article/view/3535
6. Sadeghi, J., Askarinejad, H.: Application of neural networks in evaluation of railway
track quality condition. J. Mech. Sci. Technol. 26(1), 113–122 (2012). https://doi.
org/10.1007/s12206-011-1016-5
7. Gibert, X., Patel, V.M., Chellappa, R.: Robust fastener detection for autonomous
visual railway track inspection. In: 2015 IEEE Winter Conference on Applications
of Computer Vision, pp. 694–701, January 2015
8. Sinha, D., Feroz, F.: Obstacle detection on railway tracks using vibration sensors
and signal filtering using bayesian analysis. IEEE Sens. J. 16(3), 642–649 (2016)
9. Faghih-Roohi, S., Hajizadeh, S., Núñez, A., Babuska, R., Schutter, B.D.: Deep
convolutional neural networks for detection of rail surface defects. In: 2016 Inter-
national Joint Conference on Neural Networks (IJCNN), pp. 2584–2589, July 2016
10. Marmo, R., Lombardi, L., Gagliardi, N.: Railway sign detection and classifica-
tion. In: 2006 IEEE Intelligent Transportation Systems Conference, pp. 1358–1363,
September 2006
11. Melander, M., Halme, I.: Computer vision based solution for sign detection.
Eur. Railway Rev. (2016). https://www.globalrailwayreview.com/article/30202/
computer-vision-based-solution-sign-detection/
12. Arastounia, M.: Automated recognition of railroad infrastructure in rural areas
from LIDAR data. Remote Sens. 11, 14916–14938 (2015)
13. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accu-
rate object detection and semantic segmentation. In: IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR), pp. 580–587 (2014)
14. Girshick, R.: Fast R-CNN. In: Proceedings (IEEE International Conference on
Computer Vision), vol. 2015, pp. 1440–1448 (2015)
15. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: Proceedings of the
International Conference on Computer Vision (ICCV) (2017)
16. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: unified,
real-time object detection. In: IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), vol. 2016, pp. 779–788 (2016)
17. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., Berg, A.C.:
SSD: Single shot multibox detector. In: Lecture Notes in Computer Science, vol.
9905, pp. 21–37 (2016)
18. Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P.,
Zitnick, C.L.: Microsoft COCO: common objects in context. In: Lecture Notes in
Computer Science, vol. 8693, pp. 740–755 (2014)
19. Great Britain: Department for Transport. Traffic Signs Manual. The Stationary
Office, London, United Kingdom (2013)
20. Safety, R., Board, S.: Lineside Operational Safety Signs. Railway Group Standard,
London (2009)
21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), vol.
2016, pp. 770–778 (2016)
22. Everingham, M., Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal
visual object classes (VOC) challenge. Int. J. Comput. Vis. 88(2), 303–338 (2010).
https://doi.org/10.1007/s11263-009-0275-4
23. Salton, G., McGill, M.J.: Introduction to Modern Information Retrieval. McGraw-
Hill Inc., New York (1986)
3D Conceptual Design Using Deep Learning
1 Introduction
FOOTNOTES:
[8] He represented Reading in Parliament, 1802-1820.
CHAPTER VI
BERTRAM HOUSE
Bertram House was at last finished and the beginning of 1806 saw
the Mitfords in residence. In the matter of furnishing the Doctor had
spared no expense, everything being new and of the latest pattern,
in fact the best that a fashionable London upholsterer could supply.
Of the pictures we know that the walls were well covered and that
the collection included a Gainsborough, a pair of female heads by
Greuze, and a portrait of the Doctor by Opie. We have already seen,
in Mrs. Mitford’s description of the stone-laying ceremony, that they
were attended by “the two men-servants on horseback”; this hints at
a fairly complete retinue having been installed at the Reading house,
but it was considerably augmented when the arrangements were
completed at Grazeley. Appearances counted for much in the district
and the Doctor was not the man to let slip such a grand opportunity
for ostentatious display.
His hospitality was profuse and indiscriminate, resulting in a
house-warming which extended over quite a lengthy period. As an
incentive—had he need of one—Dr. Mitford had recently been
appointed as one of the County magistrates, a tribute of appreciation
from the Whigs, of whose cause he was an earnest partizan, which
gave him an immediate rise in social status.
In time, of course, the family settled down to a more or less
ordered form of life, so ordered indeed that the Doctor created as
many excuses as possible to cover his frequent journeys to Town
and his clubs. There was sport in plenty to be had in the
neighbourhood, and of this the Doctor took full advantage, being a
familiar figure around the countryside with his gun and spaniels.
Then, too, there were the coursing meetings—the famous meetings
at Ilsley and private matches arranged between friends—none of
which were considered complete unless the Doctor were present or
his famous kennel represented. Throughout Miss Mitford’s letters,
occasionally to her father and often to friends, there are frequent
references to the greyhounds whose names, in accordance with a
custom prevalent then and still fashionable, all began with the letter
M in token of their ownership. Thus, to name only a few, we have
Mia, Manx, Marmion (a notable dog this, with an equally notable son
of the same name), Mogul, Miller, Moss-Trooper, and Mopy. For all of
these Miss Mitford ever exhibited the greatest affection, and in those
cases where a spaniel grew too old to follow the gun or a greyhound
too stiff to be matched, an asylum under Miss Mitford’s immediate
eye and care was immediately provided, and the creature was
henceforth looked upon as her own.
Taking advantage of this motherliness to dumb animals her father
frequently handed over to her some specially valuable dog from
whose later exploits as a courser he expected much. Apparently,
however, the real reason for the supposed gift was not disclosed,
with the result that when the dog was eventually removed the little
mistress gave vent to her annoyance in no measured tones.
“It is a most extraordinary thing,” she says in one communication
to her father, “that I never can have a dog that I like but you
immediately take it from me and burthen me with the care of some
detestable brute whom you in your eternal caprice fancy a good one.
Observe, however, that in giving up my own darling Mordor, I bargain
that that sulky, ungrateful, mangy beast Marmion shall be sent off as
soon as you come home, and that I shall again have my sweet
Marian to pet and comfort me.”
This was not, of course, a serious outburst, but merely the
explosion of what she doubtless considered a truly righteous
indignation, for, although she was no sportswoman, her love for her
father gave her an interest in his pursuits, and she shared with him
to the full the joy of triumph and the sorrow of defeat, while to
disparage the Mitford kennel was to offer her a personal affront. On
the other hand, she was quick to convey to the Doctor any item of
praise which she overheard or might have addressed to her. “We
called yesterday at the Fawcetts’, and the old General said he had
kept greyhounds and seen many thousands, but had never had an
idea of perfect and consummate beauty until he saw her” [a
reference, in a letter to her father, to Mia, one of the hounds].
She had a strong dislike to equestrian exercise—the rides of
babyhood across the Alresford downs with her father could not count
in this connexion—and although every inducement was offered her
to ride, an inglorious fall from a donkey quickly settled her
convictions as to her horsemanship, and her one and only riding-
habit was forthwith converted into a winter-gown. Strictly speaking,
the greater portion of her time was spent at home with her mother,
receiving visitors or lying for hours at a time on the sofa, where she
would devour a great quantity of books at a pace which, having
regard to the extraordinary knowledge she imbibed from her reading,
was truly astonishing.[9] At other times the little green chariot, their
favourite equipage, would be ordered out, calls would be returned
and the drive be possibly extended to Reading, where there would
always be plenty of shopping to do and calls to be made on the old
neighbours and friends who would have the latest news from Town
or the latest gossip of their immediate circle to retail.
With a desire to augment his income, which must have been
seriously depleted by the building operations and by the subsequent
reckless expenditure on the household, the Doctor now began to
indulge in a series of hazardous enterprises, which, with all a
gambler’s insistence, he pursued intently the while they dragged him
deeper and deeper into the mire. One of these was an extensive
speculation in coal in which he engaged with a brother of M. St.
Quintin. For this he supplied the whole of the capital in expectation of
a return of £1,500 a year, but the whole thing was a failure and, with
the exception of about £300, the capital was lost. Another
Frenchman, a man of ingenious ideas but no money wherewith to
put them to practical use, found a ready supporter in the Doctor, who
was induced to advance £5,000 on the strength of a paper scheme.
This man was the Marquis J. M. F. B. de Chabannes, and his
scheme, a supposed improved method for the lighting and heating of
houses, was embodied in a booklet which he published in 1803 with
the comprehensive title of Prospectus d’un Projet pour la
Construction de Nouvelles Maisons, Dont tous les calculs de détails
procureront une très-grande Economie, et beaucoup de
Jouissances. Unfortunately for its promoters, the scheme did not
catch on with the public, the Marquis returned to France and the
deluded Doctor continued for years to spend good money in the
hope of recovering that which was irrevocably lost by suing the
Marquis in the French courts, efforts which were all vain.
Meanwhile his fever for gambling grew apace and his absences
from home were more and more frequent and prolonged, and the
two women, being left much to themselves, conceived the notion of
arranging and copying out for the press a collection of verses
composed by the reverend father of Mrs. Mitford, Dr. Russell. They
took considerable pains with this, to which was added a special
preface by Mrs. Mitford, and when the packet was ready it was
forwarded to Dr. Mitford, in Town, with a request that he should find a
publisher and get as much as he could for it. Unfortunately, the
sanguine editors were disappointed, for no publisher sufficiently
enterprising could be found to accept the manuscript, although
sundry extracts did subsequently find a certain publicity within the
pages of the Poetical Register.
Following closely upon this effort, and in the May of 1806, Miss
Mitford went for a few days on a visit to London as the guest of
Monsieur and Madame St. Quintin, her old schoolmaster and his
wife. A short round of festivities had been arranged for her benefit,
including a visit to the Exhibition of Water Colours, evenings at the
theatres and, what appears to have been a great treat for the
impressionable Miss, some hours of two days which were spent at
Westminster Hall looking on at the trial of Lord Melville[10] and
listening to the speeches, and for which the Doctor, then in Town and
staying at Richardson’s Hotel in Covent Garden, had procured
tickets. She had now been absent from London for over three years
and, no doubt, extracted a great deal of pleasure from her visit and
its reunion with Fanny Rowden and Victoire St. Quintin, M. St.
Quintin’s sister, with both of whom, together with the Doctor, the
round of sight-seeing was enjoyed.
Mrs. Mitford stayed at home, but was kept well-posted in all the
news by the inevitable letters, full of critical details, from her dutiful
daughter. From one of these, dated from Hans Place, May 12, 1806,
we quote:—
“I have much to tell you, but it can scarcely be compressed within
the bounds of a letter. On Thursday, after I wrote, Miss Ayrton, Miss
Carp, papa, and I went to the Exhibition. There are some
uncommonly fine pictures, and it is even better worth seeing than
last year. In the evening, Victoire, Miss A. and myself went with papa
to the play to see The Provoked Husband and The Forty Thieves.
Miss Duncan in Lady Townley is most admirable. I do not much
admire Elliston as her husband. The Forty Thieves is a very
magnificent spectacle, but nothing more; for the language and music
are equally vulgar and commonplace. On Friday morning we went to
Oxford Street. I was extravagant enough to give half a guinea for a
dress skirt for myself, which I wore the next day to the trial. We were
rather disappointed in Mr. Romilly.[11] The speech in itself was
beautiful beyond description; but he wants animation, and drops his
voice at the end of every sentence.... Miss Rowden, papa, and I are
going to see Henry the Eighth to-night, and we are going to
Westminster Hall to-morrow.... I shall hope to return Thursday or
Friday; for, though I am greatly amused here, I am never quite happy
without my dear, dear mother.”
Two days later this was followed by a still more characteristic
effusion. The second day at Westminster Hall decided her that: “Mr.
Romilly is charming and interesting; but my first and greatest
favourite is Mr. Whitbread. Mr. Plumer is rather an inelegant speaker,
though very animated. I have promised papa to write some verses to
Mr. Whitbread. He has even superseded Mr. Fox in my good graces.
I did not tell you, I believe, that I had the happiness of seeing Mr. Fox
mount his horse on Saturday. I shall never again contend for his
beauty. He was obliged to lean on two people, and looked so sallow
and pale in the face, and so unwieldy in person, that I am obliged to
yield our long-disputed point.” Rather hard on poor Mr. Fox, whom,
hitherto, this exuberant young person had worshipped as a hero,
even to the extent of removing her watch-stand from the head of her
bed that it might give place to a bust of this gentleman which the
Doctor had sent from Town. On this occasion it was a case of “Off
with the old love and on with the new” in double-quick time, for,
continuing, she says: “To make me amends my new favourite is what
even you would call exquisitely handsome; a most elegant figure,
and a voice which I could listen to with transport, even if he spoke in
an unknown language. Mr. Plumer attacked him with the most
virulent irony and ridicule; and Mr. W. stood with his face turned
towards him and leant upon the desk, smiling the whole time, with
the most fascinating good humour. You know I am always an
enthusiast; but at present it is impossible to describe the admiration I
feel for this exalted character.”
We quote these extracts with no thought of ridiculing the ardent
partisan, but as a fore-shadowing of that enthusiasm and that quick
impressibility which ever seemed to dominate Miss Mitford’s life;
characteristics which often led her into excesses of transport at the
discovery, or supposed discovery, of some noble trait in the
characters of those who came within her ken, only to be as quickly
repented of; often giving unintentional pain to others and resulting in
an infinitude of trouble and annoyance to herself. Despite this
temperamental defect, however, and while her friends looked on
amazed at her infidelity, there was one to whom she remained
unwaveringly faithful to the end, though this object of her great
affection was the least worthy of all who came into her life.
FOOTNOTES:
[9] A list kept as a check on the Circulating Library account for the
years 1806 to 1811 inclusive, is a sufficient indication of this, the
number for one month alone totalling fifty-five volumes and
ranging through Fiction, Belles-Lettres, Travel and Biography.
[10] Impeached for malversation in his office as Treasurer of the
Navy. The trial lasted sixteen days. Whitbread led for the
Impeachers; Plumer—afterwards Master of the Rolls—ably
defended and secured his acquittal.
[11] Sir Samuel Romilly, Solicitor-General.
CHAPTER VII
With a proper and increasing pride in his clever daughter, the Doctor
now conceived the idea of taking her with him on an extended trip
into Northumberland, thereby affording her some acquaintance with
the scenes amidst which his family had lived for generations, a trip
which would serve the double purpose of impressing the girl with a
sense of the importance of her ancestors and present relations and
of introducing her to the latter who, although they must have heard of
her, had never yet seen her.
The journey was begun on Saturday, September 20, 1806, the first
stage, that from Reading to London being by coach. From London
they travelled in the carriage of Nathaniel Ogle, who personally
conducted them to his own place in Northumberland, from which
they were to make their various excursions in the district. Mrs.
Mitford did not accompany them, but was kept well informed, as
usual, by her daughter.
The first letter is dated from Royston, September 21: “We had a
very long interval between the parting from my most beloved darling
and the leaving Reading. The coach was completely full; and it was
fortunate papa had secured a place on the box, where he continued
during the whole journey. The company in the inside had the merit of
being tolerably quiet; and I do not remember any conversation which
lasted longer than a minute. I, certainly, ought not to complain of
their silence, as I was more than equally taciturn, and scarcely spoke
during the whole way. I was quite low-spirited, but never less
fatigued by travelling. Both Mr. Joy and Dr. Valpy[12] met us before we
left Reading, and M. St. Quintin and Victoire met us at the Bath
Hotel. As soon as Victoire left me, I retired to bed, under the idea of
pursuing our journey early in the morning. It was, however, half-past
ten before Mr. Ogle got up, and we did not leave town till twelve. We
employed the interval in going to the bookseller’s for a Cobbett, and
bought a Cary’s Itinerary, an edition of Peter Pindar, and a few plays.
The Edition of P. P. which we bought cheap, remains in town; but the
others are our travelling companions. We went by Enfield to see
Mary Ogle, and finding them at dinner we dined at Mrs. Cameron’s;
we then changed horses at Waltham Cross; again at Wade’s Mill;
and are just arrived here, where we sleep to-night. Mr. Ogle is
extremely pleasant, and the carriage very convenient. We went the
two first stages on the box of the barouche. I need not tell you, my
dearest darling, that we felt nothing so much as the loss of your
society; and I have wished myself at home fifty times in the last
twenty-four hours, to be again with my dear mamma.”
Apart from the interest which, in these days, is always attached to
an old-time account of stage-travel the letter is interesting by reason
of the variety of literature purchased for perusal on the journey. The
Cobbett referred to would probably be Cobbett’s Political Register
(then being issued in parts), and intended for the Doctor’s personal
reading; he being not only an admirer but an intimate friend of the
outspoken reformer. Cary’s Itinerary was, of course, the well-known
road-book and constant companion of all who travelled in stage-
coach days; though why Miss Mitford was not content with her
dainty, green-leather-covered copy of Bowles’ Post-Chaise
Companion in two vols.—now a valued possession of the author’s—
is difficult to understand, unless it was overlooked in the hurry and
excitement of departure. Peter Pindar’s Works, then just completed
in five vols., would be a valuable addition to the library at home, but
the purchase of the plays is significant, proving the influence which
Fanny Rowden had exercised on the mind of her pupil, inculcating a
taste for the Drama which was to be of lasting importance.
The next letter is written from Little Harle Tower—a small place
about fourteen miles from Morpeth—and is dated Sunday Evening,
September 28.—
“I arrived here with Lady Charles,[13] about two hours since, my
dearest mamma; and I find from papa that in his letter to you to-night
he never mentioned that the irregularity of the post, which never
goes oftener than three times a week from hence, will prevent our
writing again till Wednesday, when we go to Sir William Lorraine’s,
and hope to get a frank from Colonel Beaumont, whom we are to
meet there. It is only by Lord Charles going unexpectedly to Morpeth
that I am able to write this, merely to beg you not to be alarmed at
not hearing oftener. I imagine papa has told you all our plans, which
are extremely pleasant. Lord and Lady Charles stay longer in the
country on purpose to receive us, and have put off their visit to
Alnwick Castle that they may take me there, as well as to Lord
Grey’s, Colonel Beaumont’s, and half a dozen other places.”
The reference in this letter to a “frank” is one which frequently
occurs in Miss Mitford’s correspondence. It was, as Sir Rowland Hill
once said, an “expedient for saving postage”—“discreditable shifts”
another writer called them. In the days before the institution of Penny
Postage—an event which put an end to “franking”—Members
of Parliament enjoyed the privilege of having their letters delivered
and despatched free of charge. To secure this, members had merely
to write their names on the covers to ensure free passage through
the post, and frequently furnished their friends with packets of franks
which were placed aside for use as occasion required. This latter
expedient was, of course, a flagrant abuse of the privilege, and in
one year it was computed that, had postage been paid on the
franked correspondence, the revenue would have been increased by
£170,000! In an endeavour to check this abuse it was enacted that
the whole superscription must be in the handwriting of the Member,
and that the frank was only available on the date (which it was
necessary to name) which was on the cover. While the regulation
certainly diminished the quantity of franking it did not put an end to
the use of the privilege by other than Members, to whom it became
the custom to despatch an accumulated batch of letters, intended for
a number of people, with explicit instructions as to their destinations.
The annoyance caused to Members, and the general confusion
which sometimes resulted from this practice, may be better imagined
than described. Miss Mitford herself gives us an amusing account of
the troubles and trials of those who both used and abused the
franking privilege, in her sketch on “The Absent Member,” in Belford
Regis.
In the next letter which Miss Mitford wrote we have a record of
some amusing table-talk, essentially feminine in character and
which, undoubtedly, greatly impressed the observant young person
who overheard it. It is addressed still from Little Harle Tower, dated
October 3, and after a short description of the scenery, and the mud
—which caused her to beg to be excused from such excursions in
the future—she relates an account of a dinner at Sir William
Lorraine’s at which Colonel and Mrs. Beaumont were of the party.—
“Mrs. B. was so polite as to express great regret that, as she was
going from home, she could not see us at her house, but hoped,
when next we came to Northumberland, we should come to see
them at Hexham Abbey. She is a very sweet woman.... Mrs. B. told
Lady Charles that they received last year a hundred thousand
pounds from their lead mines in Yorkshire; and they never make less
than eighty thousand, independent of immense incomes from their
other estates. Mrs. B. was dressed in a lavender-coloured satin, with
Mechlin lace, long sleeves, and a most beautiful Mechlin veil. The
necklace she wore was purchased by her eldest son, a boy of
eleven, who sent it from the jeweller’s without asking the price. It is
of most beautiful amethysts; the three middle stones are an inch and
a half long and an inch wide; the price was nine hundred guineas.
Mrs. B. wished to return it; but the Colonel not only confirmed the
purchase, but gave his son some thousands to complete the set of
amethysts by a bandeau and tiara, a cestus for the waist, armlets,
bracelets, brooches, sleeve-clasps, and shoe-knots. All these she
wore, and I must confess, for a small dinner-party appeared rather
too gaily decorated, particularly as Lady Lorraine’s dress was quite
in the contrary extreme. I never saw so strong a contrast ... Colonel
Beaumont is generally supposed to be extremely weak, but I sat next
him at dinner, and he conducted himself with infinite propriety and
great attention and politeness; yet when away from Mrs. Beaumont,
he is (they say) quite foolish, and owes everything to her influence.”
Added to this cryptic description—cryptic because, read it how we
will, we cannot be sure that there is not a subtle touch of sarcasm in
the words—is a shrewd observation on another visitor whom she
calls Mr. M.
“I told you I was not enamoured of Mr. M., and I will now describe
him to you.... He is an oddity from affectation; and, I often think, no
young man affects singularity when he can distinguish himself by
anything better. He affects to despise women, yet treats them with
great respect; and he makes the most extraordinary exertions to
provoke an argument, from which he generally escapes by some
whimsical phrase.”
The letter concludes with a long list of festivities which are to take
place in honour of her visit. Following on these, they journeyed to
Kirkley, Mr. Ogle’s seat, whither it was originally intended they should
travel direct but were deterred from so doing by the hospitality
offered en route. As a matter of fact their stay at Kirkley was a short
one, due to the same cause which had prevented their earlier arrival.
The only letter addressed from Kirkley is dated Wednesday
morning, October 8:—
“We arrived here on Monday at about three o’clock; received with
great glee by the Squire, and, after taking a short walk in the garden,
returned to dress. We had some time to wait for Lord and Lady
Charles, who did not arrive before half-past five or near six, and
even then undressed. They had been detained by the axle-tree
breaking down, and the detestable roads. Without their waiting to
dress, we immediately sat down to dinner and spent a most
delightful day. In the evening we found a manuscript play which had
been sent last year for Mr. Sheridan’s perusal.[14] It is taken from a
very striking story in the Canterbury Tales, of which I have forgotten
the title.... I read it aloud to the ladies, and the gentlemen played
billiards, and occasionally visited us. The play, which bears the name
of ‘Sigendorf,’ is really extremely interesting, and much better, as to
language, than most modern productions. Sheridan had never
looked at it, and Mr. Ogle lent it to Lady Charles.
“Yesterday morning, after a long walk, Lord and Lady C. left us.
We had an excellent dinner, and amused ourselves in the evening
with the ‘Liber Veritatis,’ which is, as you may remember, a very
expensive collection of two hundred of Claude Lorraine’s sketches,
published by Boydell.
“We are going in about an hour to Little Harle ... for Mr. Ogle and
papa remain here together. We go to-morrow to Alnwick and return
the same night. To-morrow is expected to be a very full day at the
Castle, on account of the Sessions Ball. The ladies—the married
ones I mean—go in Court dresses, without hoops, and display their
diamonds and finery on the occasion.
“Mr. Ogle is quite a man of gallantry and makes his house
extremely pleasant. We talk of coming to see him again next week,
when my cousin Mary and I are left to keep house alone at Morpeth,
and my uncle and aunt go to Little Harle Tower.”
From Morpeth, on October 11, was despatched a very long letter,
too long indeed for quotation in full, but from which we must give a
few extracts. It begins:—
“In papa’s letter of yesterday, my dearest darling mamma, he
promised that I would write you a long one to-day, and I certainly
owe you one in return for the very entertaining epistle I received
yesterday. After we left Kirkley, we called at Belsay, and saw Lady
Monck and the little Atticus, who was born at Athens fifteen months
since. He is a very fine boy, very like Sir Charles. Belsay is a very old
castle, and its eccentric possessor has done all he possibly could
desire to render it still more outré by stopping up the proper road,
and obliging us to approach this fine specimen of Gothic architecture
through the farm-yard. We arrived at Little Harle to dinner; and you
would have been greatly amused at my having my hair cut by Lord
Charles’s friseur, who is by occupation a joiner, and actually
attended with an apron covered with glue, and a rule in his hand
instead of scissors. He, however, performed his office so much to my
satisfaction, that I appointed him to dress my hair the next morning
for my visit to Alnwick. While I was thus employed, Lady Swinburne
called on purpose to see me. Lady Charles said I was out walking.
She is, you know, niece to the Duke of Northumberland, and I
regretted not seeing her.