Introduction
Artificial Intelligence (AI) is experiencing enormous growth in popularity within the fields of creative arts and design. Image styletransfer based on deep neural networks
[9, 38] has become very popular both in scientific and artistic communities, influencing image creation as well as videos and movies production [14]. This approach allows achieving a decoupling of content and style in art [16], where AI has become able to generate new pieces of art using the style of one piece and the content of another.In the context of utilizing machine learning algorithms to produce new artworks, Generative Adversarial Networks (GAN) have played an integral role in the recent studies, owing to their ability to learn representations of data and to generate outputs that mimic realistic elements
[29], including songs, paintings, and sketches [3, 7, 11]. While in many GANs the input is a random noise, conditional GANs (cGANs) [25] can be used to pair a specific input with the desired generated output, as in the case of the pix2pix architecture [15], able to generate new realistic images from different representations. While most of these efforts in using GANs have focused on 2D images, the technology has become mature enough to apply these methods to 3D shapes. Multiple works have generated realistic 3D reconstruction of shapes starting from photos [35, 5] Recently, there have also been many works on generating or learning a point cloud representation of 3D objects [1, 19]. There have also been many works on converting 2D images or depthmaps to 3D [20, 31]. However, there has been little research on how to create stylistically different 3D representations of a given 3D shape.In this paper, we present Vox2Vox, a new 3D cGAN, that generalizes pix2pix to 3D. As shown in Figure 3D Topology Transformation with Generative Adversarial Networks, Vox2Vox is capable of transforming the topology and the internal structure of a 3D object, while maintaining its overall shape. We propose two volumetric topologies, the Ghirigoro and the 3D Network, and we describe how to train our Vox2Vox model to create new topology transfers. We then demonstrate the effectiveness of our approach by comparing the results obtained with our AI model against those obtained with a pure procedural algorithm.
Related Work
Before proceeding to describe the details of our approach, we will review other work on creating 3D shapes with AI.
3D Shapes Synthesis.
3D shape synthesis is used to model and reproduce shapes in 3D space. While the first applications of 3D geometry models mainly concerned video games and visual media, nowadays, reproducing 3D shapes of real world objects influences design and architecture, to selfdriving cars, to scientific and medical visualizations. Our work contributes to a shift from the manual creation of 3D models with computeraided design tools to automatically generated 3D shapes from different inputs. In addition to approaches that use probabilistic graphical models [17]
, the advent of Deep Learning brings the opportunity to automatically synthesize novel 3D shapes by
assembling parts of 3D objects, extracted from model databases, to create new compositions [13]. However, these approaches require collections of labeled 3D object parts, which despite the release of many 3D databases are still hard to acquire. To relax the labels requirement, viewbased generative models are employed. They allow for reconstruction of 3D shapes given 2D images that represent one or more view points of the object [33, 28, 32]. However, reconstructing 3D shape starting from its 2D representation often yields low quality results due to the missing information, making it hard to produce unique inputs and realistic 3D objects.Generative Adversarial Networks.
GANs have become one of the most popular neural network architectures to generate novel realistic outputs [29, 4]. GANs are composed of two neural networks, the Discriminator, trained to distinguish between the real and the generated inputs, and the Generator, trained to produce new outputs that look real. GANs have been used to generate voxelbased object representations in the 3D space [5, 35], starting from random noises. However, producing a desired shape requires exploring the latent space in order to find the correct random input for the Generator. In this context, cGANs introduce nonrandom inputs to control the outputs of the Generator [10]. cGANs have been employed as 3D shape Generators in [35, 36]
, where 2D images and a conditional probability density function are respectively used to condition the output of the Generator and map the input to the desired 3D shapes. However, only few works exist about controlling the output of the Generator with 3D shapes as conditional inputs.
[27] presents a 3DcGAN able to perform rotations of volumes in the 3D space. In comparison, our approach aims to transfer the shape of the input, while modifying its volumetric topology to obtain a novel collection of 3D objects, and, to the best of our knowledge, our is the first 3DcGAN model for 3D topology transformations.Style Transfer.
Style transfer aims to learn the content and the structure of an input element while changing its style based on a different style source. In the past, many works have been released that perform 2Dto2D style transfer on images and videos [16, 9]. 2Dto3D style transfer is mainly employed to apply specific textures or colors from 2D samples to 3D objects [26]. [18] presents a novel 2Dto3D style transfer approach able to perform gradientbased 3D mesh editing operations to modify also the surface of the 3D shapes based on the image used as style source. [23] is one of the first works about 3Dto3D style transfer: the proposed algorithm computes the analogy between one source element and the related target and applies it to synthesize new outputs based on different sources. This deformationbased approach is also used to generate different poses of animal meshes [34] and modify the design of 3D furniture, buildings models, and different classes of objects [37, 21, 22, 24]. However, these works require a specific formulation of analogy between the different parts of the analyzed objects that limits their application to few collections of 3D models. In this context, our work lays the foundation for a novel 3Dto3D style transfer. Indeed, with our 3DcGAN model for 3D topology transformations, we can say that the style is sedimented in the trained network and multiple styles are supported with different weights of our Vox2Vox Generator. A future step would be allowing both shapes and styles as inputs of the Generator.
Approach
The goal of this paper is to convert a 3D shape into an alternative 3D representation inspired by the original. We will focus on converting the shape into a 3D Network, which is an abstract representation in the same vein as cubism in art. A network (graph) consists of two types of components: 1) Nodes, being the entities to be connected and 2) Links, which are the wires connecting the nodes. Like scaffold that supports the interior of the 3D object, we want the nodes to be placed in suitable locations inside the shape and links to connect these nodes to form a 3D Network.
However, as we will discuss later, a random or a uniform choice of where to position nodes and link points doesn’t result in aesthetically pleasing results. Thus, we have to find a better way to choose where to place the nodes and link points. This requires us to find what part of the 3D shape should contain more nodes in order for the shape to be represented better than random nodes, and we would have to handcode the criterion for finding this node density. If, however, the aesthetics is changed slightly, e.g. having curvy links like a scribble taking the shape of the network, we will need to, again, handcode a procedure for the new aesthetics. Solving this inverseproblem of finding an algorithm or rule for suitable distribution of node and link positions inside the 3D shape can be very difficult. In contrast, training a system that would learn by examples could be much easier. Therefore, we want our Vox2Vox cGAN to be able to produce the node and link distributions, using voxels. Once we have this distribution, we can use a procedural algorithm, we called Procedural Network, that takes the voxels as an input (together with the number of nodes and the number of points to be used for links connecting nodes) and generates the final 3D Network representation.
In this way the main challenge moves from creating an algorithm that performs the transformation of the desired topology (in this case Network 3D) to creating the dataset necessary for training the AI model. Fortunately, creating arbitrary 3D Networks and converting them to spacefilling 3D blobs is much easier than doing the reverse, as we will discuss in the section about model training. Once we have 3D Networks with their respective 3D figures, the 3DcGAN is able to learn how to convert the latest to the former, while finding a pure procedural algorithm for doing so may be difficult and an optimization process may be costly. Another benefit of this approach is that, if we want to choose a different aesthetics, we can create a new aestheticsrelated training dataset and not worry about the inverse problem.
Proposed Method
As stated in the previous section, solving the problem of creating a neural network to perform topology transfer is equivalent to creating a neural network that can take a 3D shape and produce a voxelsbased distribution of nodes and links to put inside it. Thus, the output should preserve the general 3D structure of the input. It was shown in [15] that when preserving the overall location of features in the input and output is important, the UNet architecture performs much better than an encoderdecoder architecture, owing to its skip layers. UNet [30]
is also known to converge with relatively little training data, making it ideal for our application. We, therefore, choose a UNet architecture for the main component of our neural network. However, given a 3D shape, there are many different ways that one can fill this shape with a network. This means that when generating 3D blobs from networks, many different network layouts may end up having similar 3D blobs. This will confuse a neural network, as the same input is being assigned different 3D Networks as “label”. This is precisely where a cGAN becomes useful. Unlike a simple neural network, where multiple labels would result in the network choosing the average of those labels, in a GAN the loss is minimized as long as the GAN learns to produce one of the correct labels, so that the Discriminator cannot tell the output from the real data. In summary, we need a cGAN which has UNet architecture for its Generator and a Discriminator suitable for classifying 3D shapes. This is essentially exactly what the pix2pix architecture
[15] does, only in 3D instead of 2D. Thus, the architecture of Vox2Vox will be very similar to pix2pix, but the 2D convolutional layer will be replaced with 3D convolutions and the number of layers and filters will be different.Pipeline Overview.
Figure 1 shows an overview of the pipeline for converting a 3D shape into a 3D Network. First we transform a 3D input mesh to a filled voxel representation, given a target 3D resolution. The filled voxels are passed to the Vox2Vox Generator which outputs two channels of voxels: ch. 1 for the distribution of where to put nodes and ch. 2 for distribution of points along links connecting the nodes. Finally, the voxel distributions for nodes and links are passed to the Procedural Network (described in detail below) algorithm to produce the final 3D Network.
Model Architecture.
Generator 


Filters  Params  

Input Layer    
Down3D*  2,080  
Down3D  131,392  
Down3D  524,928  
Down3D  1,049,216  
Up3D  1,049,216  
Up3D  1,048,896  
Up3D  262,304  
Up3D*  8,194 
Discriminator 

Filters  Params  

Input Layer    
Down3D*  3,088  
Down3D  32,800  
Down3D  131,136  
Down3D*  32,769 
* No Batch Normalization
Kernel size=8Table 1 shows the details of the modules used in the Vox2Vox Generator and Discriminator. The Generator has a UNet architecture, with four Down3D modules in the encoder and four Up3D modules in the decoder each of which contains a 3D convolutional layer. During training, the input shape is and the encoder layers have and filters, encoding the input to shape. The decoder layers have and filters. is the number of desired output channels, which for the 3D Network problem is two, one for node and one for link distribution. The Up3D modules also get the output of the encoding layers as input, as in UNet. The Discriminator consists of four 3D convolutional layers, yielding a output. The cost function for the Discriminator decides whether on all the patches of the node and link distributions match the input 3D blob in a certain way. If the Generator successfully produces a node and link distribution which matches the blob over all the similar to how the real data matches the blob, the Discriminator will get fooled. Because of the third dimension, the training process of Vox2Vox is both very memory intensive and computationally very expensive. That is why we chose to do the training in the resolution of voxels. However, during prediction we increase the input size to produce higher quality outputs.
Indeed, a nice feature of UNet is that, since all layers are convolutional, we can produce larger outputs by simply modifying the input shape of the trained network. Note that, this increase in input size will not result in larger “features”. For example, if the Vox2Vox trained at resolution produces nodes with radius voxels maximum, increasing the input shape to will still only produce nodes of radius . Similarly, if at resolution the maximum length of links produced is voxels, it will be the same when the input size is changed to . This is because the convolutional layers in all layers will still have the same receptive field of, say on the input, at both and resolutions. However, what this allowS us to do is that we can feed arbitrary large inputs to a trained Vox2Vox ( was the maximum size our GPU memory allowed) and convert them into 3D Networks. Moreover, making the input larger allows us to convert objects with much higher details than the size of the objects in the training set. Therefore, after we train Vox2Vox on voxels, we change the input shape to and produce higher quality outputs. We used Vox2Vox on a variety of different 3D sculptures and passed the output to the Procedural Network Algorithm to extract the generated 3D Network (Fig. 3).
Model Training
The training procedure of Vox2Vox (Figure 3) is similar to the pix2pix one: we present the Discriminator with a pair of inputs, which is the concatenation of the array of the 3D shape with the array of real or fake output along the channels, yielding a array. For the real pair, the label is a array of ones, and for the fake pair the array is zero. The main difference is our training data is 3D and completely synthetic (see sections about training data below). Some points to note is that, we exploited the convolutional nature of the UNet layers also during training. During the training process, the input is a binary array (0 in empty spaces and 1 inside the 3D shape) and output of the Generator is a array, where is the number of channels. While this training converges to reasonably good results after around 10 epochs, we found that first training on
binary arrays and very small networks, and then doing transfer learning by increasing the input resolution resulted in faster convergence to reasonably good results.
3D Network Training Data.
We create a variety of networks first and then lay them out in 3D using a custom forcedirected layout algorithm, inspired from [6], which makes sure nodes do not overlap. We then create two arrays from the network. One is a array of the 3D shape of the layout, merging nodes and link segments by replacing them with overlapping large and small spheres, respectively. This 3D shape is the input of the Generator. The second is a array, with the first channel being the nodes and the second the links, which are again replaced by spheres, this time with smaller radii (Figure 3). Then, we replace nodes with spheres and segments along links with spheres which overlap and make the connected links. We generate about 700 such networks and heavily augment the dataset by rotating them in multiples of 20 degrees about the x,y and z axes to create a dataset of about 30,000 data points. We trained Vox2Vox with a few different thicknesses for links and sizes for nodes to determine a good choice for the sizes of the nodes and links so as to avoid spacefilling links and nodes, which would let the Generator exploit this structure and fool the discriminator by simply filling the inside of the shape. Figure 2 shows a sequence of 3D Network outputs from the Vox2Vox Generators while growing of epochs.
Note on Input Network Topology.
There exist many different generative processes for producing networks which result in very different connectivity patterns. We found that using random graphs [8] resulted in poor results. In these graphs, any node is equally likely to connect to any other node, resulting in a completely random network, with all nodes having a similar degree (i.e. number of links attached to them). The Vox2Vox Generator was able to produce reasonable positions for the nodes, but failed at producing good links between them. Our hypothesis is that, since links in these networks are completely random, the Discriminator overfitted to the few samples it had been trained on and was never satisfied with any other variation of the links. In contrast, when we trained the Vox2Vox on networks generated using a “rich gets richer” (BarabásiAlbert (BA)) model [2] the results improved dramatically. In the BA model some nodes, known as hubs, have significantly more links than other nodes. The BA model has “hubs” which are nodes with a much higher degree than most other nodes, which contrasts it strongly from the random ER network where all nodes have more or less the same degree. We chose the size of the nodes in the 3D array to be a function of their degree. When making the voxel representation, we assign lager node radii nodes with a higher degree. We believe that this predictable correlation between the node size and the density of links was learned by the Generator and was exploited to fool the Discriminator, allowing it to produce good results.
Usually, to train GANs a single input is presented to the network at a time [10], meaning batch size 1. This will force the Generator to try to learn the features of a single example. However, this will result in a very slow progression of the training. On the other hand, larger batch size yields faster convergence, but presenting a large batch of data result in more of an averaged output, stopping it from choosing a single pattern. Therefore, in the initial stages of the training, we set the batch size to 8 to allow for a faster approach towards good filters. In later stages, we reduced the batch size to 2 and then 1 to finetune the results.
Ghirigoro: Training on a Second Topology. Aside from converting shapes to 3D Networks, we also tried a second, related topology. The “Ghirigoro” (doodle) topology consisted of converting a 3D shape into a long scribble that mimics its shape. This style has only one output channel, which is a set of long curvy lines, crossing more rarely than the 3D Network case, and there are no nodes. To construct the final doodle shape from the GAN we used the Procedural Network algorithm with different settings. The result is not exactly a doodle, as the Procedural Network may crosslink two parallel pieces of the doodle, resulting in a network. Nevertheless, the aesthetic is close to a doodle. The pipeline is shown at the bottom of Figure 1 and the results are presented along with the 3D Network case in Figure 5.
Procedural Network algorithm
We interpret the two output channels of the Vox2Vox Generator as node and link distributions. To extract a network from these distributions, we first use Kmeans to choose centers for nodes and for points along the links. We choose a higher K for link points than nodes in the Kmeans. We then form a network from points that are close to each other, connecting link points and node centers in a dense network. To avoid connecting points that are far from each other, we partition the space into small regions and connect points within each partition to each other. To avoid artifacts from the position of partition walls, we repartition the space multiple times by slightly shifting the partition walls by a random value. This method of connecting nodes and link points works better than a knearest neighbor method, as all nodes need not have the same degree, and hence neighbors. This results in a series of connected paths connecting nodes using via intermediate link points. Finally, we use the A* search algorithm
[12] to extract links between nodes.To see how using the Vox2Vox output as a prior for node and link distribution compares against not using having this prior, we apply the above procedure to the original 3D shape voxels. This means that, to find node locations we just take the filled voxels of the 3D shape and apply Kmeans to it. This will fill the interior of the shape with randomly chosen node locations. Our default was choosing 300 node locations. We then do the same thing for link points, except that we choose 50 times more points to serve as points along links (i.e. points). The rest of the procedure is the same. Comparing the outcomes, we observe that the baseline procedure results in curved and broken links, while the proposed Vox2Vox model produces 3D shapes with more effectively distributed links and nodes.
Results
Figure 4 shows the results of the proposed 3D topology style transfer approach. For each 3D input model, we present its volumetric transformations into the 3D Network topology and the Ghirigoro topology. We then compare in Figure 5 the results of the baseline pure procedural algorithm that applies the volumetric topology transformation directly on the 3D input.
As is evident in Figure 5, the baseline (fourth column, black model) looks quite messy. The baseline also requires a much higher node density to be able to capture the shape of the object. In comparison, using the Vox2Vox output a prior for the node and link distributions of the 3D Network (second column, red) results in much nicer results with much fewer nodes an links. Vox2Vox chooses node locations very economically and strategically, preserving the overall structure of the shape with much fewer nodes. Vox2Vox also does a very interesting job with the Ghirigoro style (third column, blue). The results all preserve the shape of the object very well, while using very few curved lines.
Thus, our results show that it is possible to perform complex 3D topology style transfer using our Vox2Vox architecture, if we possess a good number of samples in the desired style, or are able to generate such samples. The samples could have any arbitrary or random shape and the task of Vox2Vox is to figure out how to convert a given shape into that style. Moreover, the resolution of the output can be modified, as long as the size of the features does not need to scale with the input size. Lastly, one could use other procedures to convert the Vox2Vox output to high quality 3D shapes, like our Procedural Network algorithm.
Conclusion and Future Direction
In this paper, we presented a novel 3Dto3D topology transfer paradigm based on transformations in 3D space. In particular, we built a 3D conditional GAN, Vox2Vox, that performs volumetric transformations to modify the internal structure of any 3D object, while maintaining its overall shape. We described our complete pipeline to apply our approach to two different topologies: the 3D Network and the Ghirigoro. The results obtained by employing our methodology are novel and inspirational. We compared the outputs of the pipeline while using or not the 3DcGAN and found that using the Vox2Vox output as a prior distribution results in much nicer outcomes where features are placed in strategic positions in the 3D shape preserving its structural features. As a future direction, we plan to improve the 3Dto3D topology transfer by given also the topology as a conditional input of the generative network. To do that, the machine learning algorithm has to learn itself the abstraction of the topology from a given 3D object.
References
 [1] (2017) Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392. Cited by: Introduction.
 [2] (1999) Emergence of scaling in random networks. science 286 (5439), pp. 509–512. Cited by: Note on Input Network Topology..
 [3] (2017) Deep learning techniques for music generationa survey. arXiv preprint arXiv:1709.01620. Cited by: Introduction.
 [4] (2018) Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096. Cited by: Generative Adversarial Networks..

[5]
(2016)
Generative and discriminative voxel modeling with convolutional neural networks
. arXiv preprint arXiv:1608.04236. Cited by: Introduction, Generative Adversarial Networks..  [6] (2018) Structural transition in physical networks. Nature (to appear). Cited by: 3D Network Training Data..
 [7] (2017) A learned representation for artistic style. Proc. of ICLR. Cited by: Introduction.
 [8] (1960) On the evolution of random graphs. Publ. Math. Inst. Hung. Acad. Sci 5 (1), pp. 17–60. Cited by: Note on Input Network Topology..

[9]
(2016)
Image style transfer using convolutional neural networks.
In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
, pp. 2414–2423. Cited by: Introduction, Style Transfer..  [10] (2014) Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680. Cited by: Generative Adversarial Networks., Note on Input Network Topology..
 [11] (2017) A neural representation of sketch drawings. arXiv preprint arXiv:1704.03477. Cited by: Introduction.

[12]
(1968)
A formal basis for the heuristic determination of minimum cost paths
. IEEE transactions on Systems Science and Cybernetics 4 (2), pp. 100–107. Cited by: Procedural Network algorithm.  [13] (2015) Analysis and synthesis of 3d shape families via deeplearned generative models of surfaces. In Computer Graphics Forum, Vol. 34, pp. 25–38. Cited by: 3D Shapes Synthesis..
 [14] (2017) Realtime neural style transfer for videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7044–7052. Cited by: Introduction.

[15]
(2017)
Imagetoimage translation with conditional adversarial networks
. arXiv preprint. Cited by: Introduction, Proposed Method. 
[16]
(2016)
Perceptual losses for realtime style transfer and superresolution
. In European Conference on Computer Vision, pp. 694–711. Cited by: Introduction, Style Transfer..  [17] (2012) A probabilistic model for componentbased shape synthesis. ACM Transactions on Graphics (TOG) 31 (4), pp. 55. Cited by: 3D Shapes Synthesis..
 [18] (2018) Neural 3d mesh renderer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3907–3916. Cited by: Style Transfer..
 [19] (2018) Point cloud gan. arXiv preprint arXiv:1810.05795. Cited by: Introduction.
 [20] (2019) Synthesizing 3d shapes from silhouette image collections using multiprojection generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5535–5544. Cited by: Introduction.
 [21] (2015) Style compatibility for 3d furniture models. ACM Transactions on Graphics (TOG) 34 (4), pp. 85. Cited by: Style Transfer..
 [22] (2015) Elements of style: learning perceptual shape style similarity. ACM Transactions on Graphics (TOG) 34 (4), pp. 84. Cited by: Style Transfer..
 [23] (2014) Analogydriven 3d style transfer. In Computer Graphics Forum, Vol. 33, pp. 175–184. Cited by: Style Transfer..
 [24] (2018) Towards 3d neural style transfer.. In AIIDE Workshops, Cited by: Style Transfer..
 [25] (2014) Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784. Cited by: Introduction.
 [26] (2012) 3D material style transfer. In Computer Graphics Forum, Vol. 31, pp. 431–438. Cited by: Style Transfer..
 [27] (2018) Paired 3d model generation with conditional generative adversarial networks. arXiv preprint arXiv:1808.03082. Cited by: Generative Adversarial Networks..
 [28] (2016) Volumetric and multiview cnns for object classification on 3d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5648–5656. Cited by: 3D Shapes Synthesis..
 [29] (2015) Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434. Cited by: Introduction, Generative Adversarial Networks..
 [30] (2015) Unet: convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computerassisted intervention, pp. 234–241. Cited by: Proposed Method.
 [31] (2018) Pixels, voxels, and views: a study of shape representations for single view 3d object shape prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3061–3069. Cited by: Introduction.
 [32] (2017) Synthesizing 3d shapes via modeling multiview depth maps and silhouettes with deep generative networks. In The IEEE conference on computer vision and pattern recognition (CVPR), Vol. 3, pp. 4. Cited by: 3D Shapes Synthesis..
 [33] (2015) Multiview convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pp. 945–953. Cited by: 3D Shapes Synthesis..
 [34] (2004) Deformation transfer for triangle meshes. In ACM Transactions on Graphics (TOG), Vol. 23, pp. 399–405. Cited by: Style Transfer..
 [35] (2016) Learning a probabilistic latent space of object shapes via 3d generativeadversarial modeling. In Advances in Neural Information Processing Systems, pp. 82–90. Cited by: Introduction, Generative Adversarial Networks..
 [36] (2018) Learning descriptor networks for 3d shape synthesis and analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 8629–8638. Cited by: Generative Adversarial Networks..
 [37] (2010) Stylecontent separation by anisotropic part scales. In ACM Transactions on Graphics (TOG), Vol. 29, pp. 184. Cited by: Style Transfer..
 [38] (2017) Unpaired imagetoimage translation using cycleconsistent adversarial networkss. In Computer Vision (ICCV), 2017 IEEE International Conference on, Cited by: Introduction.
Comments
There are no comments yet.