Обучение нейросетей компьютерного зрения в видеоиграх

SHAFAEI et al.: USING VIDEO GAMES TO TRAIN COMPUTER VISION MODELS 1
Play and Learn: Using Video Games to
Train Computer Vision Models
Alireza Shafaei
http://cs.ubc.ca/~shafaei
James J. Little
http://cs.ubc.ca/~little
Mark Schmidt
http://cs.ubc.ca/~schmidtm
Department of Computer Science
University of British Columbia
Vancouver, Canada
Abstract
Video games are a compelling source of annotated data as they can readily pro-
vide fine-grained groundtruth for diverse tasks. However, it is not clear whether
the synthetically generated data has enough resemblance to the real-world images to
improve the performance of computer vision models in practice. We present experi-
ments assessing the effectiveness on real-world data of systems trained on synthetic
RGB images that are extracted from a video game. We collected over 60,000 syn-
thetic samples from a modern video game with similar conditions to the real-world
CamVid and Cityscapes datasets. We provide several experiments to demonstrate
that the synthetically generated RGB images can be used to improve the performance
of deep neural networks on both image segmentation and depth estimation. These re-
sults show that a convolutional network trained on synthetic data achieves a similar
test error to a network that is trained on real-world data for dense image classifica-
tion. Furthermore, the synthetically generated RGB images can provide similar or
better results compared to the real-world datasets if a simple domain adaptation tech-
nique is applied. Our results suggest that collaboration with game developers for an
accessible interface to gather data is potentially a fruitful direction for future work in
computer vision.
1 Introduction
Deep neural networks have been setting new records in almost all of the computer vision
challenges and continue to grow in the broader field of artificial intelligence. One of the
key components of a successful deep network recipe is the availability of a sufficiently
large dataset for training and evaluation. The Imagenet Large Scale Visual Recognition
Challenge (ILSVRC) [35], with over 1 million training images, has been one of the pri-
mary benchmarks for evaluation of several deep network architectures over the past few
years. As architectures such as Resnet [15] approach a 3.57% top-5 error on the Imagenet
classification task, it is likely that a larger and more complex dataset will become nec-
essary soon. However, compiling datasets has its difficulties – depending on the desired
granularity of the groundtruth, the costs and the labor could grow rather quickly.
c 2016. The copyright of this document resides with its authors.
It may be distributed unchanged freely in print or electronic forms.
arXiv:1608.01745v2[cs.CV]15Aug2016

2 SHAFAEI et al.: USING VIDEO GAMES TO TRAIN COMPUTER VISION MODELS
Meanwhile, to cope with limited available data, there are several approaches to the train-
ing of deep networks. The common practices include data augmentation through various
image transformations [19] or use of pre-trained networks that are trained for a similar task
as a warm-start in the original optimization problem. A more radical approach is to synthe-
size training samples through computer-generated (CG) imagery.
One of the main concerns in the use of CG data for the training is whether the source
domain is close enough to the target domain (i.e., the domain of real-world images) to make
an effective application possible. For depth images of the Kinect sensors, for instance, sev-
eral lines of work [33, 38, 40, 45] demonstrate the efficacy and practical similarities between
computer-generated depth images and the output of consumer level depth sensors. There
are also methods that rely on indirect representations such as the HOG templates [16] or
geometrical descriptors [23, 42], that can be efficiently synthesized and used. For the direct
synthesis and application of RGB data, however, there is a limited amount of work that only
investigates simple rendering of 3d CAD models [3, 4, 22, 30, 43] or synthetic text [14]. The
question of interest is whether existing video games with photorealistic environments can be
a useful source of synthetic RGB data to address computer vision tasks.
Although video games generate images from a finite set of textures, there is variation in
viewpoint, illumination, weather, and level of detail which could provide valuable augmen-
tation of the data. In addition to full control over the environment, video games can also
provide us with groundtruth data such as dense image class annotations, depth information,
radiance, irradiance, and reflectance which may not be straightforward, or even possible, to
collect from real data. Other measures, such as the precise location of the game character
within the environment, could be useful for development of visual SLAM algorithms.
In this work, we focus our attention on the RGB domain of a modern video game and
run various experiments to gauge the efficacy of using CG RGB data directly for computer
vision problems. We collect over 60,000 outdoor images under conditions similar to the
CamVid [5] and the Cityscapes [7] datasets and present experiments on two computer vision
problems: (i) dense image annotation, and (ii) depth estimation from RGB. We show that a
convolutional network trained on synthetic data achieves a similar test error to a network that
is trained on real-world data. Furthermore, after fine-tuning, our results show a network that
is pre-trained on synthetic data can outperform a network that is pre-trained on real-world
data.
2 Previous Work
Synthetic Data. Synthetic data has a successful history in computer vision. Taylor et al. [44]
present a system called ObjectVideo Virtual Video (OVVV) based on Half-life [1] for eval-
uation of tracking in surveillance systems. Marin et al. [27] extend OVVV to perform
pedestrian detection with HOG [8] features. In the recent literature, a variety of meth-
ods [4, 22, 30, 43] use 3d CAD models with simple rendering to tackle vision problems.
Peng et al. [30], and Sun and Saenko [43] use non-photorealistic 3d CAD models to improve
object detection. Lim et al. [22], and Aubry et al. [4] use CAD models for detection and
object alignment in the image. Aubry and Russell [3] use synthetic RGB images rendered
from CAD models to analyze the response pattern and the behavior of neurons in the com-
monly used deep convolutional networks. Rematas et al. [31] use 3d models to synthesize
novel viewpoints of objects in real world images. Stark et al. [42], Lim et al. [23], and
Liebelt and Schmid [21] learn intermediate geometric descriptors from 3d models to per-

form object detection. Butler et al. [6] present the synthetic Sintel dataset for evaluation of
optical flow methods. Synthetic depth images are also successfully used for human pose
estimation [38, 40] and hand pose estimation [33, 45]. Kaneva et al. [17] study robustness of
image features under viewpoint and illumination changes in a photorealistic virtual world. In
contrast to previous work, we take a different approach to synthetic models. Instead of ren-
dering simple 3d CAD models in isolation, we take a step further and collect synthetic data
in a simulated photorealistic world within the broad context of street scenes. We are specifi-
cally targeting the use of modern video games to generate densely annotated groundtruth to
train computer vision models.
Dense Image Classification. Deep Convolutional Networks are extensively used for
dense image segmentation [26, 29, 48, 49]. The fully convolutional networks of Long et
al. [26] is among the first to popularize deep network architectures that densely label input
images. Zheng et al. [49] build on top of the architecture in Long et al. [26] and integrate
CRFs with Gaussian pairwise potentials to yield a significant gain in image segmentation.
The current state-of-the-art methods, such as the one presented by Liu et al. [25], use variants
of fully convolutional networks as a building block on top of which different recurrent neural
networks or graphical models are proposed. We use the basic fully convolutional architecture
of Long et al. [26] as it provides the basis of the follow-up developments in this area.
Depth Estimation from RGB. One of the early studies on unconstrained depth estima-
tion from single RGB images is the work of Saxena et al. [36] in which the authors present
a hierarchical Markov random field to estimate the depth. More recently, Zhuo et al. [50]
present an energy minimization problem that incorporates semantic information at multiple
levels of abstraction to generate a depth estimate. Li et al. [20] and Liu et al. [24] use deep
networks equipped with a conditional random field that estimates the depth. More recently,
Eigen and Fergus [10] presented a multi-scale deep convolutional architecture that can pre-
dict depth, normals, and dense labels. Instead of regressing against the metric data directly,
Zoran et al. [51] propose a general method to estimate reflectance, shading, and depth by
learning a model to predict ordinal relationships. The input image is first segmented into
SLIC superpixels [2] on top of which a multi-scale neighbourhood graph is constructed. Zo-
ran et al. [51] use the neighbourhood graph and generate ordinal queries on its edges, and
then use the results to construct a globally consistent ranking by solving a quadratic program.
We apply the method of Zoran et al. [51] in our study and show improvement in the depth
estimation task through the use of synthetic data.
Transfer Learning. A closely related area of work to our study is transfer learning
(see Pan and Yang [28] for a review). The early studies of transfer learning with Deep
Convolutional Networks successfully demonstrated domain adaptation through pre-training
on source data and fine-tuning on target data [9, 39, 47]. Further studies such as the work of
Ganin et al. [12] present more sophisticated approaches to domain adaptation. In this work,
we apply the most widely used fine-tuning approach to domain adaptation and leave further
studies on feature transferability of the synthetic data to the future work.
3 Datasets
In this work, we focus on outdoor on-the-road datasets as we can easily collect this data
from the video game. For real-world images, there are a few available datasets [5, 7, 13, 37]
of which we use Cityscapes [7] and CamVid [5]. The selected datasets provide the most
appropriate setting regarding the available groundtruth and visual similarity.

Figure 1: Densely labeled samples from the VG+ dataset. The label space of this dataset is
the same as the CamVid [5] dataset.
3.1 VG Dataset
The dataset consists of over 60,000 frames collected from a video game1. To gather this
data we use a camera on the hood of a car, similar to the configuration of the CamVid [5]
or Cityscapes [7] datasets. The weather is kept fixed at sunny and the time of the day is
fixed at 11:00 AM. This atmospheric setting was chosen to make the synthetic data similar
to the real-world data, although note that a key advantage of VG is that it would be easy
to sample data under non-ideal conditions (while it might be impossible to collect reliable
real data under many conditions). The image resolution of the game is 1024 × 768 with
the highest possible graphics configuration. The autonomous driver randomly drives around
the city while obeying the traffic laws. Every second, a sample data is collected from the
game. Each sample contains the RGB image, groundtruth semantic segmentation, depth
image, and the surface normals. The groundtruth semantic segmentation that we were able
to automatically extract from the game is over the label set {Sky, Pedestrian, Cars, Trees}.
The help of the game developers is likely to be required in order to extract more labels.
We also consider a label-augmented version of the VG dataset, which we call VG+. To
augment the label space with additional classes, we use SegNet [18], one of the top perform-
ing methods on CamVid [5] dataset with available code and data, to classify the images of
the VG. We then refine the labels with the true labels of the VG dataset and clean-up the data
automatically using the depth and the surface normals. See Fig. 1 for samples.
Note that, unlike the groundtruth dense annotations of VG, the groundtruth label space
Y in VG+ is noisy and may partially exhibit the biases of the SegNet [18]. However, the
input space X remains the same. Consequently, the results that are derived from VG+ are
weaker but still can provide us with useful insight as the input synthetic RGB is intact. In
principle, it should be possible to automatically collect the groundtruth dense annotation of
all the objects in the game.
3.2 Real-world Datasets
CamVid [5]. The CamVid dataset contains 701 densely labelled images collected by a
driving car in a city (see Fig. 2). The label set is {Sky, Building, Pole, Road Marking, Road,
Pavement, Tree, Sign Symbol, Fence, Vehicle, Pedestrian, Bike}. The data is split into 367,
101, and 233 images for train, validation, and test.
1Due to legal concerns we were asked to not disclose the title of the game.

Figure 2: Two random images and the corresponding annotations from the CamVid [5]
dataset.
Figure 3: Two random images with the corresponding dense annotations from the
Cityscapes [7] dataset.
Cityscapes [7]. The Cityscapes dataset offers 5000 finely annotated images and 20,000
coarsely annotated images of urban street scenes over 33 labels collected in cities in Ger-
many and Switzerland. In addition to the dense and coarse image annotations, the dataset
also includes car odometry readings, GPS readings, and the disparity maps that are calcu-
lated using the stereo camera. See Fig. 3 for pixel-level annotation examples. The densely
annotated images are split into sets of 2975, 500, and 1525 for training, validation, and test.
With the exception of ‘Road Marking’, all the other 11 classes of CamVid are present in
Cityscapes. The groundtruth for the test set is kept private for evaluation purposes. Thus, we
perform the Cityscapes evaluations on the validation set.
4 Dense Image Classification
For this task we use the fully convolutional network (FCN) of Long et al. [26]. More specif-
ically, we use the FCN8 architecture on top of a 16-layer VGG Net [41]. FCN8 is a simple
architecture that only uses convolution and deconvolution to perform classification, and pro-
vides a competitive baseline for dense image classification. All of the networks will be
trained with the same settings: SGD with momentum (0.9), and pre-defined step sizes of
10−4, 10−5, and 10−6 for 50, 25, and 5 epochs. We use the MatConvNet [46] library.
To align our datasets with respect to the label space, we define two subsets of the data.
The datasets CamVid, Cityscapes, and VG refer to a variation in which the label space is
limited to {Pedestrian, Cars, Trees, Sky, Background}. In this setting, the label space of the
VG is precise and automatically extracted from the game. The second variation CamVid+,
Cityscapes+, and VG+ refers to the setting in which the label space is the full 12 classes
of the original CamVid dataset and the labels in the synthetic VG+ is noisy. When we are
evaluating on Cityscapes+ we omit the missing ‘Road Marking’ class.

Sky
B
uilding
PoleR
oad
M
arking
R
oad
Pavem
ent
TreeSign
Sym
bol
Fence
C
arPedestrian
B
icyclist
A
verage
0.2
0.4
0.6
0.8
1 0.90
0.91
0.91
0.91
0.69
0.68
0.73
0.73
0.53
0.58
0.58
0.60
0.83
0.84
0.87
0.88
0.81
0.81
0.79
0.80
0.84
0.84
0.94
0.94
0.80
0.83
0.80
0.82
0.45
0.51
0.65
0.57
0.65
0.79
0.82
0.84
0.85
0.91
0.91
0.91
0.75
0.83
0.82
0.85
0.68
0.72
0.70
0.75
0.73
0.77
0.79
0.80
Class Accuracy on CamVid+
Baseline
Real
Synthetic
Mixed
Figure 4: The per-class accuracy on the test set of CamVid+ dataset. The baseline is
trained on the target dataset, the real is pre-trained on the real alternative dataset, and the
synthetic is pre-trained on VG+. The Mixed approach is pre-trained on both synthetic
and the real alternative dataset. Pre-training the baseline with the synthetic VG+ improves the
average accuracy by 6%, while pre-training with the real-world Cityscapes+ improves
the average by 4%.
CamVid+ Cityscapes+
Model Pixel Acc. Class Acc. Mean IoU Pixel Acc. Class Acc. Mean IoU
Baseline 79% 73% 47% 83% 77% 50%
Real 80% 77% 51% 83% 77% 50%
Synthetic 82% 79% 52% 84% 79% 51%
Mixed 82% 80% 53% 84% 79% 52%
Table 1: Evaluation of different pre-training strategies on CamVid+ and Cityscapes+,
comparing pixel accuracy, mean class accuracy, and mean intersection over union (IoU).
Pre-training on synthetic data consistently outperforms the equivalent model that is only
pre-trained on real-world data. The mixed pre-training strategy gives the most improvement.
4.1 Evaluation with Fine-tuning
To measure the effect of using synthetic data, we look at the performance in a domain adap-
tation setting in which we perform fine-tuning on the target dataset. In this approach we suc-
cessively train our dense classifier on different datasets and then examine the performance.
In the first experiment we focus on the CamVid+ dataset. We train four different networks
and evaluate their performance on CamVid+ to analyze the influence of our synthetic data.
Each model is pre-trained on an alternative dataset(s) before training on the target dataset.
For instance, the experiment with name Synthetic means that the network has been pre-
trained on the synthetic VG dataset first, and then fine-tuned on the target dataset (see Fig. 4).
The Real counterpart is the network that is pre-trained on the alternative real-world dataset
first. The Mixed approach is when we pre-train on both synthetic and real-world data first.
Figure 4 compares our training strategies with respect to the per-class accuracy. Pre-
training on VG+ improves the average accuracy more than pre-training on Cityscapes+
does, 79% vs. 77%. The most improvement is for the class ‘Sign Symbol’ where pre-training
with VG+ improves the accuracy by 20%. The highest improvement is achieved when we
pre-train on both real and synthetic datasets. Table 1 shows a summary of our results.

0 20 40 60 80
Epoch
0.6
0.8
1
1.2
Loss
Objective Value (Cross-Entropy Loss)
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.6
0.65
0.7
0.75
0.8
0.85
Accuracy
Class Average Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
Figure 5: The influence of various pre-training approaches on the Cityscapes+. The left
image is the objective function throughout training, and the right image is the class average
accuracy. The solid lines are the evaluation results on the training set, the dashed lines are
the results on the validation set. Pre-training on synthetic data gives a better initialization
and final local minima compared to pre-training on a real-world dataset. The figure is best
viewed in color.
We can also analyze the benefits of using this synthetic data from an optimization per-
spective. Figure 5 shows the objective value and the class average accuracy of Cityscapes+
during the last optimization stage. Pre-training on the VG+ datasets yields better results on
train and validation in comparison to both the baseline, and the pre-trained version on the
real-world CamVid+ dataset. Pre-training on the VG+ data provides a better initialization
and final local minima in the training procedure. In both of the results, the synthetic data
provides a clear improvement over the baseline, and helps as much or more than pre-training
on the real data. More figures and analysis of these experiments can be found in the supple-
mentary material.
4.2 Cross-dataset Evaluation
We also evaluate the models in a cross-dataset setting in which the network is evaluated on
a dataset other than the one it was trained on. We train three dense image classifiers on
each dataset and examine the cross-dataset accuracy of these classifiers. The purpose of this
experiment is to see how much the domain of our synthetic data differs from a real-world
dataset in comparison to another real-world dataset.
Figure 6 shows the cross-dataset accuracy on CamVid and Cityscapes. The first
observation, as anticipated, is that the domain of a real-world dataset is more similar to the
domain of another real-world dataset on average. Even though the VG network has been only
trained on synthetic data, in ‘pedestrian’, ‘car’, or ‘trees’, it competes or even in two cases
outperforms the network trained on real data. Although the VG network does not outperform
the real counterpart on average, the small gap indicates that the network with synthetic data
has learned relevant features and is not overfitting to the game specific textures that can be
an obstacle to generalization to the real-world domain.
Figure 7 shows the per-class accuracy for each network on Cityscapes+. Similar to
the previous results, the domain of CamVid+ is more similar to the Cityscapes+ than the
synthetic VG+ dataset is. However, VG+ gives better results for pole, tree, sign symbol, fence,
and bicyclist. On average, the network that is trained on CamVid+ gives a 60% accuracy,
and the network obtained from synthetic data gives a similar 56% accuracy. The similarity
in performance suggests that for the training of computer vision models the synthetic data

Backg. Sky Car Pedes. Trees Average
0.2
0.4
0.6
0.8
1 0.89
0.89
0.87
0.94
0.87
0.68
0.83
0.87
0.87
0.90
0.91
0.67
0.89
0.79
0.46
0.89
0.86
0.71
Class Accuracy on CamVid
Baseline
Real
Synthetic
0.2
0.4
0.6
0.8
1
0.88
0.92
0.71
0.98
0.88
0.59
0.89
0.76
0.73
0.92
0.59
0.64
0.93
0.71
0.74
0.92
0.77
0.68
Class Accuracy on Cityscapes
Baseline
Real
Synthetic
Figure 6: The cross-dataset per-class accuracy on CamVid (left) and Cityscapes (right).
The baseline is trained on the target dataset, the real is trained on the real alternative dataset,
and the synthetic is trained on VG.
Sky
B
uilding
Pole
R
oad
Pavem
ent
TreeSign
Sym
bol
Fence
C
ar
Pedestrian
B
icyclist
A
verage
0.2
0.4
0.6
0.8
1
0.96
0.87
0.66
0.71
0.69
0.60
0.55
0.38
0.39
0.92
0.91
0.53
0.82
0.70
0.56
0.82
0.64
0.76
0.72
0.27
0.36
0.70
0.51
0.64
0.83
0.69
0.57
0.77
0.46
0.39
0.72
0.43
0.64
0.77
0.60
0.56
Class Accuracy on Cityscapes+
Baseline
Real
Synthetic
Figure 7: The cross-dataset per-class accuracy on the validation set of the Cityscapes+
dataset. The baseline is trained on Cityscapes+, the real is trained on the CamVid+, and
the synthetic is trained on VG+.
provides a reasonable proxy to the real world images. For more figures on this experiment
please see the supplementary material.
5 Depth Estimation from RGB
The Cityscapes dataset also provides the disparity images which we will be using in this
setting. For this problem we use the method of Zoran et al. [51], in which the underlying
deep network is queried with two points in the image and has to decide which one is closer
to the camera. This method is more attractive than the other techniques because it only relies
on the ordinal relationships and not on the measurement unit in the target dataset. This is
useful for our experiments because the depth images of the video game are tailored to im-
prove the visualization effects in the rendering pipeline and are not directly comparable to
the measurement units of the real-world datasets. We use the same deep network architecture
as Zoran et al. [51]. The images are first decomposed into superpixels. The center of each
superpixel is a node in the graph. Adjacent superpixels are connected on this graph in a mul-
tiscale fashion. The two end-points of each edge are then used to query a deep network that
classifies the relative depth of two input patches as {=,>,<}. A global ranking of the pixels
is then generated based on these queries by solving a quadratic program. Depending on the
target ordinal relationships we can estimate the depth, shading, or reflectance. We apply this

Epoch
0 20 40 60 80
Loss
0.55
0.6
0.65
0.7
0.75
0.8
Baseline-train
Baseline-valid
Synthetic-train
Synthetic-valid
Epoch
0 20 40 60 80
Accuracy
0.66
0.68
0.7
0.72
0.74
0.76
0.78
Classification Accuracy
Baseline-train
Baseline-valid
Synthetic-train
Synthetic-valid
Figure 8: The influence of pre-training on synthetic data for the depth estimation task on the
proposed network architecture of Zoran et al. [51].
method for depth estimation and train the underlying deep network on the Cityscapes and
the VG datasets. While the networks in the previous experiments focused on higher level
visual cues and abstractions, the nature of the network in this problem is concerned with the
mid-level visual cues, which provides further insight to the quality of the synthetic data.
In Figure 8 we look at the influence of pre-training the network on the VG dataset vs.
directly learning from the real dataset Cityscapes. Similar to the previous experiments
we observe that pre-training with the synthetic data gives us a consisent improvement in
initialization and the final local minima, in both validation and training. The final patch
classification accuracy on the validation set is improved from 75% to 76%. For more figures
on this experiment see the supplementary material.
6 Conclusion
Concurrently, a number of independent studies that explore similar ideas have been published
recently. Gaidon et al. [11] present the Virtual KITTI dataset and show experiments on multi-
object tracking tasks. Ros et al. [34] present the SYNTHIA dataset of urban scenes and also
demonstrate improvement in dense image segmentation using synthetic data. Our study on
using video games complements the recent work by providing analysis on the photorealistic
output of a state-of-the-art game engine. Furthermore, an independent study to be published
by Richter et al. [32] provide a similar analysis of using video games which we invite the
reader to review for a complete picture.
As video games progress towards photorealistic environments, we can also use them for
the training of computer vision models at no extra cost. We delivered a proof of concept
by exploring the use of synthetic RGB images that we extracted from a video game. Our
approach goes beyond the use of simplistic 3d CAD models as we collect a synthetic dataset
that encompasses the broad context of street scenes. We presented several experiments to
compare our synthetic dataset with the existing real-world ones. Our experiments show
that in a cross-dataset setting, the deep neural networks that we trained on synthetic RGB
images have a similar generalization power as the networks that we trained on real-world
data. Furthermore, with a simple domain adaptation technique such as fine-tuning, pre-
training on synthetic data consistently yielded better results than pre-training on real data.
Moreover, we showed that pre-training on synthetic data resulted in a better initialization and
final local minima in the optimization. On a network that classifies the ordinal relationship
of image patches, we also observed how pre-training on synthetic data leads to improvement
in optimization and final performance.

Our results suggest that RGB images collected from video games with photorealistic
environments are potentially useful for a variety of computer vision tasks. Video games
can offer an alternative way to compile large datasets for direct training or augmenting real-
world datasets. Although we limited our attention to a fixed atmospheric setting to facilitate
comparing to the existing real-world datasets, a key advantage is the ability to capture and
use data under a variety of environmental parameters. For example, it would be easy to
collect data under winter, night, fog, and other conditions where reliable real data is difficult
to obtain.
Acknowledgments. We would like to thank Ankur Gupta for helpful comments and discus-
sions. We gratefully acknowledge the support of NVIDIA Corporation through the donation
of the GPUs used for this research. This work was supported in part by NSERC under Grant
CRDPJ 434659-12.
References
[1] Half-Life 2. http://www.valvesoftware.com/games/hl2.html.
[2] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and
Sabine Susstrunk. SLIC superpixels compared to state-of-the-art superpixel methods.
TPAMI, 2012.
[3] Mathieu Aubry and Bryan C. Russell. Understanding deep features with computer-
generated imagery. In ICCV, 2015.
[4] Mathieu Aubry, Daniel Maturana, Alexei A. Efros, Bryan C. Russell, and Josef Sivic.
Seeing 3d chairs: Exemplar part-based 2d-3d alignment using a large dataset of cad
models. In CVPR, 2014.
[5] Gabriel J. Brostow, Julien Fauqueur, and Roberto Cipolla. Semantic object classes in
video: A high-definition ground truth database. Pattern Recognition Letters, 2009.
[6] Daniel J. Butler, Jonas Wulff, Garrett B. Stanley, and Michael J. Black. A naturalistic
open source movie for optical flow evaluation. In ECCV, 2012.
[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler,
Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes
dataset for semantic urban scene understanding. In CVPR, 2016.
[8] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection.
In CVPR, 2005.
[9] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng,
and Trevor Darrell. DeCAF: A deep convolutional activation feature for generic visual
recognition. In ICML, 2014.
[10] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels
with a common multi-scale convolutional architecture. In ICCV, 2015.
[11] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy
for multi-object tracking analysis. In CVPR, 2016.

[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle,
François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial train-
ing of neural networks. JMLR, 2016.
[13] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets
robotics: The KITTI dataset. The International Journal of Robotics Research, 2013.
[14] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text locali-
sation in natural images. arXiv preprint arXiv:1604.06646, 2016.
[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for
image recognition. In CVPR, 2016.
[16] Mohsen Hejrati and Deva Ramanan. Analysis by synthesis: 3d object recognition by
object reconstruction. In CVPR, 2014.
[17] Biliana Kaneva, Antonio Torralba, and William T. Freeman. Evaluation of image fea-
tures using a photorealistic virtual world. In ICCV, 2011.
[18] Alex Kendall, Vijay Badrinarayanan, and Roberto Cipolla. Bayesian segnet: Model un-
certainty in deep convolutional encoder-decoder architectures for scene understanding.
arXiv preprint arXiv:1511.02680, 2015.
[19] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classiﬁcation with
deep convolutional neural networks. In NIPS, 2012.
[20] Bo Li, Chunhua Shen, Yuchao Dai, Anton van den Hengel, and Mingyi He. Depth and
surface normal estimation from monocular images using regression on deep features
and hierarchical CRFs. In CVPR, 2015.
[21] Joerg Liebelt and Cordelia Schmid. Multi-view object class detection with a 3d geo-
metric model. In CVPR, 2010.
[22] Jasmine J. Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing ikea objects: Fine
pose estimation. In ICCV, 2013.
[23] Joseph J. Lim, Aditya Khosla, and Antonio Torralba. FPM: Fine pose parts-based
model with 3d CAD models. In ECCV, 2014.
[24] Fayao Liu, Chunhua Shen, and Guosheng Lin. Deep convolutional neural ﬁelds for
depth estimation from a single image. In CVPR, 2015.
[25] Ziwei Liu, Xiaoxiao Li, Ping Luo, Chen-Change Loy, and Xiaoou Tang. Semantic
image segmentation via deep parsing network. In ICCV, 2015.
[26] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for
semantic segmentation. In CVPR, 2015.
[27] Javier Marin, David Vázquez, David Gerónimo, and Antonio M. López. Learning
appearance in virtual scenarios for pedestrian detection. In CVPR, 2010.
[28] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions
on Knowledge and Data Engineering, 2010.

[29] George Papandreou, Liang Chen, Kevin Murphy, and Alan L. Yuille. Weakly- and
semi-supervised learning of a deep convolutional network for semantic image segmen-
tation. In ICCV, 2015.
[30] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object
detectors from 3d models. In ICCV, 2015.
[31] Konstantinos Rematas, Tobias Ritschel, Mario Fritz, and Tinne Tuytelaars. Image-
based synthesis and re-synthesis of viewpoints guided by 3d models. In CVPR, 2014.
[32] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data:
Ground truth from computer games. In ECCV, 2016.
[33] Grégory Rogez, James S. Supancic III, and Deva Ramanan. First-person pose recogni-
tion using egocentric workspaces. In CVPR, 2015.
[34] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M.
Lopez. The SYNTHIA dataset: A large collection of synthetic images for semantic
segmentation of urban scenes. In CVPR, 2016.
[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma,
Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C.
Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015.
[36] Ashutosh Saxena, Sung H. Chung, and Andrew Y. Ng. 3-D depth reconstruction from
a single still image. IJCV, 2008.
[37] Timo Scharwächter, Markus Enzweiler, Uwe Franke, and Stefan Roth. Efﬁcient multi-
cue scene segmentation. In Pattern Recognition. 2013.
[38] Alireza Shafaei and James J. Little. Real-time human motion capture with multiple
depth cameras. In CRV, 2016.
[39] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn
features off-the-shelf: an astounding baseline for recognition. In CVPR Workshops,
2014.
[40] Jamie Shotton, Ross Girshick, Andrew Fitzgibbon, Toby Sharp, Mat Cook, Mark
Finocchio, Richard Moore, Pushmeet Kohli, Antonio Criminisi, and Alex Kipman.
Efﬁcient human pose estimation from single depth images. TPAMI, 2013.
[41] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-
scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[42] Michael Stark, Michael Goesele, and Bernt Schiele. Back to the future: learning shape
models from 3d CAD data. In BMVC, 2010.
[43] Baochen Sun and Kate Saenko. From virtual to reality: Fast adaptation of virtual object
detectors to real domains. In BMVC, 2014.
[44] Geoffrey R. Taylor, Andrew J. Chosak, and Paul C. Brewer. OVVV: Using virtual
worlds to design and evaluate surveillance systems. In CVPR, 2007.

[45] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous
pose recovery of human hands using convolutional networks. TOG, 2014.
[46] Andrea Vedaldi and Karel Lenc. MatConvNet – Convolutional neural networks for
MATLAB. In Proceeding of the ACM Int. Conf. on Multimedia, 2015.
[47] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are
features in deep neural networks? In NIPS. 2014.
[48] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions.
In ICLR, 2016.
[49] Shuai Zheng, Sadeep Jayasumana, Bernardino Romera-Paredes, Vibhav Vineet,
Zhizhong Su, Dalong Du, Chang Huang, and Philip Torr. Conditional random ﬁelds as
recurrent neural networks. In ICCV, 2015.
[50] Wei Zhuo, Mathieu Salzmann, Xuming He, and Miaomiao Liu. Indoor scene structure
analysis for single image depth estimation. In CVPR, 2015.
[51] Daniel Zoran, Phillip Isola, Dilip Krishnan, and William T. Freeman. Learning ordinal
relationships for mid-level vision. In ICCV, 2015.

0 20 40 60 80
Epoch
100
Loss
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Accuracy
Pixel Classification Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.3
0.4
0.5
0.6
0.7
0.8
IoU
Mean IoU
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.6
0.7
0.8
0.9
1
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
Figure 9: The influence of various pre-training approaches on the CamVid dataset. The
solid lines are the evaluation results on the training set, the dashed lines are the results on the
validation set.
Supplementary Material
Evaluation with Fine-tuning
Figures 9-12 compare the behavior of our convolutional networks in the training phase.
We present the evolution of objective value, pixel classification accuracy, class average
accuracy, and mean intersection-over-union for CamVid, Cityscapes, CamVid+, and
Cityscapes+ datasets. Pre-training on synthetic data consistently improves the initial-
ization and the final solution, and in most cases also outperforms pre-training on real-world
data.
Figure 13 compares the per-class accuracy of each training strategy on the test set of
CamVid and the validation set of Cityscapes. Using synthetic data yields a consis-
tent improvement over the baseline. On CamVid, pre-training on real data leads to a bet-
ter model than pre-training on synthetic data, but the mixed approach has the best accu-
racy. On Cityscapes, however, pre-training on synthetic data has a higher average ac-
curacy than pre-training on real-world data. Figure 14 shows the per-class accuracy on the
Cityscapes+ dataset. Similar to the previous experiments using synthetic data results in
more improvement than using real-world data. Combining synthetic and real data gives the
highest performance boost in these experiments.
Cross-dataset Evaluation
In the cross-dataset setting, we train one network on each dataset and evaluate the accuracy
of each network on the other datasets. The purpose of this experiment is to measure and

0 20 40 60 80
Epoch
0.2
0.3
0.4
0.5
0.6
Loss
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.82
0.84
0.86
0.88
0.9
0.92
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.5
0.55
0.6
0.65
0.7
IoU
Mean IoU
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.84
0.86
0.88
0.9
0.92
0.94
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
Figure 10: The inﬂuence of various pre-training approaches on the Cityscapes dataset.
The solid lines are the evaluation results on the training set, the dashed lines are the results
on the validation set.
0 20 40 60 80
Epoch
100
Loss
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.4
0.5
0.6
0.7
0.8
0.9
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.1
0.2
0.3
0.4
0.5
0.6
IoU
Mean IoU
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
Figure 11: The inﬂuence of various pre-training approaches on the CamVid+ dataset. The
solid lines are the evaluation results on the training set, the dashed lines are the results on the
validation set.

0 20 40 60 80
Epoch
0.6
0.8
1
1.2
Loss
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.7
0.75
0.8
0.85
0.9
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.35
0.4
0.45
0.5
0.55
IoU
Mean IoU
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
0 20 40 60 80
Epoch
0.6
0.65
0.7
0.75
0.8
0.85
Accuracy
Baseline-train
Baseline-valid
Real-train
Real-valid
Synthetic-train
Synthetic-valid
Mixed-train
Mixed-valid
Figure 12: The inﬂuence of various pre-training approaches on the Cityscapes+ dataset.
The solid lines are the evaluation results on the training set, the dashed lines are the results
on the validation set.
0.6
0.8
1
Class Accuracy on CamVid
0.89
0.90
0.89
0.90
0.94
0.94
0.94
0.94
0.83
0.91
0.90
0.91
0.89
0.93
0.93
0.94
0.89
0.92
0.91
0.93
0.89
0.92
0.91
0.92
Baseline
Real
Synthetic
Mixed
0.6
0.8
1
Class Accuracy on Cityscapes
0.88
0.88
0.88
0.88
0.98
0.98
0.98
0.98
0.89
0.90
0.92
0.92
0.92
0.91
0.93
0.93
0.93
0.93
0.93
0.93
0.92
0.92
0.93
0.93
Baseline
Real
Synthetic
Mixed
Figure 13: The per-class accuracy on CamVid (left) and Cityscapes (right).
Sky
B
uilding
Pole
R
oad
Pavem
ent
TreeSign
Sym
bol
Fence
C
ar
Pedestrian
B
icyclist
A
verage
0.2
0.4
0.6
0.8
1
Class Accuracy on Cityscapes+
0.96
0.97
0.97
0.97
0.71
0.70
0.72
0.72
0.55
0.56
0.58
0.59
0.92
0.92
0.93
0.93
0.82
0.83
0.84
0.84
0.81
0.81
0.81
0.81
0.72
0.71
0.72
0.72
0.70
0.69
0.73
0.73
0.83
0.83
0.86
0.86
0.77
0.77
0.79
0.80
0.68
0.67
0.72
0.69
0.77
0.77
0.79
0.79
Baseline
Real
Synthetic
Mixed
Figure 14: The per-class accuracy on Cityscapes+.

Sky
B
uilding
Pole
R
oad
Pavem
ent
TreeSign
Sym
bol
Fence
C
ar
Pedestrian
B
icyclist
A
verage
0
0.2
0.4
0.6
0.8
1
Class Accuracy on CamVid+
0.90
0.81
0.65
0.69
0.59
0.64
0.53
0.60
0.69
0.92
0.90
0.76
0.85
0.88
0.87
0.80
0.66
0.47
0.44
0.70
0.68
0.65
0.77
0.85
0.85
0.91
0.89
0.75
0.77
0.51
0.66
0.21
0.19
0.73
0.71
0.65
Baseline
Real
Synthetic
Figure 15: Cross-dataset evaluation. The per-class accuracy on the test set of the CamVid+
dataset.
compare the generalization power of the networks that are trained on synthetic or real data
only. Figure 15 shows the per-class accuracy for evaluation on the Camvid+ dataset. The
Baseline network is directly trained on the target dataset, while the Real network is
trained on the alternative real dataset, and the Synthetic network is trained on synthetic
data only. Without domain adaptation, both of the Real and Synthetic networks have a
lower accuracy than the Baseline. The network that is trained on real data has a better ac-
curacy than the network that is trained on synthetic data only. Even though the Synthetic
network is only trained on synthetic data, it outperforms the real network on ‘Building’,
‘Pole’, and ‘Fence’. While the Synthetic network does not exceed the accuracy of the
Real network on average, the small gap indicates that the network with synthetic data is
relying on relevant features and is not merely overfitting to the game specific textures.
6.1 Depth Estimation from RGB
Zoran et al. [51] present a depth estimation method that only relies on the ordinal rela-
tionships between a set of image patch pairs. The image is first decomposed into SLIC [2]
superpixels. A deep convolutional network classifies the ordinal relationship between the ad-
jacent superpixels by generating a local relationship label {<, =, >} with the corresponding
probabilities. A quadratic program is then constructed to generate a total ordering (ranking)
over the superpixels which will represent the depth. Note that this method does not rely
on the depth measurement unit. Hence, it is not directly comparable to the prior work that
directly regress to the depth value. We use this method because the depth information that
is collected in the video game is not directly comparable to the real-world depth metrics.
The ordinal relationships, however, can be consistently inferred from the extracted depth in-
formation. In the main paper, we demonstrated how using the synthetic RGB images can
improve the patch classifier of Zoran et al. [51]. Figure 16 shows the groundtruth depth and
the predicted depth image of a sample input from the Cityscapes dataset.

Figure 16: (top left) A sample image from the Cityscapes [7] dataset, (top right) decom-
position of the RGB image to SLIC superpixels [2], (bottom left) the groundtruth disparity
map, (bottom right) the globalized depth output of the method presented by Zoran et al. [51].

Обучение нейросетей компьютерного зрения в видеоиграх

More Related Content

What's hot (19)

Similar to Обучение нейросетей компьютерного зрения в видеоиграх (20)

More from Anatol Alizar (20)

Recently uploaded (20)

Обучение нейросетей компьютерного зрения в видеоиграх