Improved Adversarial Systems for 3D Object Generation and Reconstruction
This paper describes a new approach for training generative adversarial networks (GAN) to understand the detailed 3D shape of objects. While GANs have been used in this domain previously, they are notoriously hard to train, especially for the complex joint data distribution over 3D objects of many categories and orientations. Our method extends previous work by employing the Wasserstein distance normalized with gradient penalization as a training objective. This enables improved generation from the joint object shape distribution. Our system can also reconstruct 3D shape from 2D images and perform shape completion from occluded 2.5D range scans. We achieve notable quantitative improvements in comparison to existing baselines.
|Edward J. Smith|
|Department of Computer Science|
|McGill University Canada|
|Department of Computer Science|
|McGill University Canada|
noticebox[b]1st Conference on Robot Learning (CoRL 2017), Mountain View, United States.\end@float
Keywords: Model learning, Multimodal perception, Robotic vision
Understanding the 3D shape of objects in the world is a crucial step for many areas of robotics. Across object categories, shape is used for classification and to comprehend affordances. Within categories, fine shape details and textures contribute to successful manipulation. The complexity of the combined shape distribution has limited the success of parametric models historically used for the task. Recently, techniques for unsupervised learning of deep networks trained on large data-sets of synthetic 3D shapes have produced models capable of representing the detailed shape space [1, 2]. This allows the generation of strikingly realistic sample shapes for simulation or visualization as well as the interpretation of robot sensory data.
This paper improves and extends a recent method known as 3DGAN , which has shown the ability to generate realistic samples of 3D object shapes, to categorize 3D shapes, and to reconstruct 3D shape from 2D images. 3DGAN is based upon the original generative adversarial network (GAN)  architecture and training approach, which is well-known to suffer from instability. While 3DGAN is proficient at producing high quality objects from single classes, we have observed that it is difficult to train on distributions involving multiple distinct object classes in varied poses, as can be seen in Figure 1, leading the authors to primarily report results from independent models, each trained only on a single object category. We have been motivated to pursue joint training over a mixture of categories, following the goal of completely data-driven modeling. Can this complex shape distribution be captured without category label supervision?
The GAN architecture assigns each object it has experienced to a point in a latent space. The remainder of this space can allow generation of novel realistic objects, but only if the GAN training has converged successfully to the data distribution. For simple data distributions, such as objects from a single category, a simple network structure such as linear interpolation may suffice. Complex data distributions, such as those formed by the sharp transitions between diverse object categories, lead to harder training problems.
In order to capture these wider and more complicated distributions, we propose the use of the recently proposed Wasserstein training objective with gradient penalty, a technique that has been successful in training GANs for other complex tasks. We call our resulting shape model the 3D Improved Wasserstein Generative Adversarial Network (3D-IWGAN). Our results show that this approach improves the stability of training for complex distributions. We are able to generate multiple object categories from multiple viewpoints using a single joint shape network. We show that 3D-IWGAN can be applied to reconstructing 3D shape from 2D images by integrating our training with a Variational Auto-Encoder (VAE) . This leads to our new 3D-VAE-IWGAN system which achieves state of the art accuracy for the task, achieving an mean average precision score of 61.7 percent on the IKEA dataset, an 8.6 percent increase. This extended system relies on novel update rules to synchronize between encoding, generation and discrimination tasks, which stabilized our model’s learning and were key in establishing their convergence. Our 3D-VAE-IWGAN system also allows us to perform shape completion, where full 3D shape is determined based on a 2.5D range scan, effectively removing the occlusion due to the sensor’s geometry. In order to ensure reproducible experimental comparison, our code for all of these systems is publicly available on a GitHub repository111https://github.com/EdwardSmith1884/3D-IWGAN.
2 Related Works
The pioneering work of Roberts  and Marr  exemplify early shape models composed of 3D geometric primitives with recognizable 2D properties. Similarly intuitive properties are still in common use, such as symmetric parts , skeletons  and CAD wire-frames .
Recently, generating distributions based upon the parameters of a complex function learned from data have been utilized to represent shapes without the need for any prior structural knowledge . Recent examples include: GP-LVMs , Convolutional Neural Networks , Recurrent Neural Networks , Deep Belief Networks , Generative Adversarial Networks [16, 3] and Deep Convolutional Auto-encoders . The work that we extend upon in this paper, 3DGAN , falls in this category. Several of these shape models have been shown in practical applications such as interactive editing  and reconstructing room geometry [19, 9, 20] and depth  from a single image can which act as import cues for understanding scenes at a higher level [22, 23].
As shown in Figure 2, our 3D-IWGAN architecture utilizes voxelized objects from a data set to train deep generator and discriminator networks in tandem, with the goal of generating realistic 3D object shapes. Our discussion will focus on the improvements we made to the GAN training process that were crucial in capturing this complex distribution.
3.1 3D Generation
The GAN algorithm is a generative system consisting of two networks, the generator and the discriminator, which learn together to mimic a target distribution . The generator converts normally distributed latent vectors into samples which the discriminator then classifies as either generated or real. They play the following minimax game with value function :
where is sampled from the target and is a normally distributed vector. Theoretical results have shown  that GAN training minimizes the Kullback–Leibler (KL) divergence between the data and generated distributions, with convergence to equality in idealized conditions. When training deep networks in this fashion, the use of gradient descent leads to unstable learning. One must carefully balance the learning of the generator and the discriminator, or gradients can vanish which prevents improvement. The Wasserstein GAN algorithm (WGAN)  attempts to fix these flaws by instead minimizing the Wasserstein distance between the distributions. This metric is defined for , the set of all joint distributions whose marginal distributions are and , as
The WGAN scheme has been shown to produced far more stable convergence, and allow one to track convergence through the discriminator’s loss . The WGAN system’s main departure from the original system is clipping the discriminators weights to lie within a compact space, which forces the the discriminator to lie within the set of 1-Lipschitz functions. This constraint is a key in ensuring constructed Wasserstein distance is always continuous and almost always differentiable.
The new Improved Wasserstein GAN training system (IWGAN)  provides a more desirable method for enforcing the Lipschitz constraint. As opposed to weight clipping, it penalizes deviation of the discriminator’s gradients from unity, as the gradients of a differentiable function are at most 1 if and only if it is a 1-Lipschitz function. This results in the following loss function for the discriminator:
where and are the generator and target distribution, and is the distribution sampling uniformly on a straight line between and . Gulrajani et al.  shows that weight clipping pushes the networks towards overly simple functions and can often lead to vanishing or exploding gradients. When penalizing the gradient, however, these problems no longer occur, leading to more complex network functions and more stable convergence. We build upon the ideas leveraged in the 3D-GAN system , applying our own new architectures and substituting the standard GAN training model for the new IWGAN system.
In our new 3D-IWGAN system, shortcomings of the original system, including a poor ability to generalize to more varied object distributions and unstable convergence are ameliorated. The gradient penalty specified above is added to the discriminator’s loss, with . Objects at random points on straight lines between our generated objects and real objects are used for its calculation, as suggested in . As opposed to the original GAN scheme , the generator now learns only every 5 batches, while the discriminator learns on every batch, which leads to more stable convergence. Batch normalization layers are not present in our discriminator as they cause the discriminator’s output for each sample to be dependent upon the values of all samples in the observed batch, and so corrupts our gradient penalty. In addition, no activation function is applied to the discriminator’s final output, and Adam  optimization is used to train both the generator and the discriminator, both with learning rate . These changes, which in combination with the network architectures below we refer to as 3D-IWGAN, lead to more stable convergence and allow for quantitative tracking of our models’ convergence.
3.2 3D Object Reconstruction from Single Images
We tackle the problem of reconstructing 3D objects from RGB images by combining our new 3D-IWGAN system with the VAE-GAN system , in a novel training scheme. A VAE  is generative, encoder-decoder system for learning complex distributions. The encoder observes samples from the target distribution and produces a vector of means and variances parameterizing a set of Gaussians, which are sampled to produced a latent vector. This vector is passed to the decoder which attempts to reproduce the original sample. This encoding system is performed to facilitate generation by sampling the latent vector from a normal distribution. Both networks are penalized on the sample’s reconstruction error, while the encoder’s loss also includes a regularizing KL divergence term encouraging the Gaussians to resemble standard normal distributions. The VAE-GAN system  is an algorithm to combine the efficient encoding of VAE’s with the generative power of the GAN system to produce samples that can be conditioned on prior data . It works by leveraging a single network for both the generator and decoder, and combining their loss functions.
As is typical for VAE-GAN systems , our VAE’s decoder network is simultaneously used for the GAN’s generator network. The VAE’s encoder converts an image into a 400 dimensional vector of means of variances, which are sampled using Gaussians to produce our latent vector. The latent vector is then passed through the decoder/generator network to generate a reconstructed object, which is then given to the discriminator to pass judgment on its validity. The loss functions for the encoder network and the decoder/generator network are then:
respectively, where x is the target sample, is the generated sample (generated from an encoded image in the first equation, and a random latent vector in the second) , and are the means and variances produced by the encoder, and . The discriminator’s loss remains the same.
The same generator and discriminator networks as used for 3D generation are implemented into this system and the encoder network is a simple 5 layer convolution neural network. During training, the discriminator and encoder networks are trained at every batch while the generator only learns every 5 batches. This last point is key to the integration of the systems, as if the encoder is not trained alongside the discriminator at every iteration the system will not converge. This makes sense as both networks should learn similar features about the objects being created at approximately the same rate. We refer to this combined system as 3D-VAE-IWGAN, and a summary diagram of this system can be viewed in Figure 2. This application is chosen to demonstrate our model’s improved generative power, and to introduce our novel system for integrating the Wasserstein Algorithms with the VAE-GAN system. With our new system we achieve state of the art reconstruction results on the IKEA Dataset .
3.3 Reconstruction of 3D Objects from Single Perspective Depth Scans
We accurately reconstruct an object’s complete 3D shape and volume when presented with only a depth map of the object from a single persecutive, as one might receive from a single snapshot of a Kinect scan. We tackle this problem to show the generative power of our 3D-VAE-IWGAN system, to highlight that our models are applicable to realistic robotic problems, and to demonstrate that our system is easily applicable to tasks involving reproducing 3D shapes from multiple input types.
For 3D generation, the ModelNet10 Dataset  was used, consisting of roughly 57000, 32x32x32 resolution objects, from 10 object classes in 12 evenly spaced orientations. The 10 classes are bathtubs, beds, chairs, desks, dressers, monitors, nightstands, sofas, tables and toilets. The goal of our 3D generation system is to be trained on the entire dataset and successfully produce varied objects from all of these classes and orientations simultaneously.
An entirely synthetic dataset, that we constructed, was exclusively used to train our models for 3D object reconstruction from single images. This dataset consisted of the 6 object classes from the ShapeNet Dataset , rendered in front of backgrounds images from the SUN dataset  at random poses, lighting conditions, and distances, and overlayed with random textures from the Describable Textures Dataset . Accompanying each image was its ground truth object from the ShapeNet Dataset, at 20x20x20 voxel resolution. The IKEA Dataset  was used to test our models, which consists of a set of 759 images picturing a large array of objects, and the corresponding object models. These objects fall into 6 categories: beds, bookcases, chairs, desks, sofas, and tables, and are evaluated at resolution 20x20x20. The dataset presents a strong evaluation tool for heavily occluded images in realistic scenes, using only the constraint that the object is centered within the image.
We constructed a second synthetic dataset, taking 10 random perspectives for each object in the ModelNet10 dataset in order to train models for the shape completion task. A test set of entirely unseen objects was held back for evaluation, examples of which, can be observed in the first rows of Figure 8 and Figure 9. In order to ensure the models trained on our synthetic data was sufficient for application to real world data, a small set of Kinect depth maps sampled from the Large Dataset of Object Scans  was converted into voxel representation and also used for evaluation.
3.5 Network Architectures
New network architectures were designed for the generator and the discriminator in our systems to allow for quicker training times, the integration of new GAN techniques, and to conform to the requirements of the IWGAN algorithm. The input to the generator, a 200 dimensional normally distributed latent vector, is passed though a fully connected layer with 2048 nodes, followed by four 3D deconvolutional layers with stride length 2 and kernel size 4. Each deconvolutional layer is followed by a batch normalization layer and a ReLU activation layer, apart from the final layer which is only followed by a tanh activation layer. The output of the network is a 32x32x32 resolution voxel grid.
The discriminator’s network input, a 32x32x32 resolution object, is passed through 4 3D convolutional layers, flattened, and then condensed to a single value through a final fully connected layer. Each convolutional layer posses stride length 2 and kernel size 4, and is followed by a Leaky ReLU activation layer. The network architectures for these two systems can be seen in Figure 3. The bases for these two architectures are the DC-GAN networks, due to their success in generating high quality images . The combination of new training system under the IWGAN algorithm and new architectures described above results in our novel generative model which we refer to as as 3D-IWGAN. Figure 2 summarizes this system.
In order to generate objects at 20x20x20 resolution (needed to test on the IKEA Dataset), slightly downscaled versions of the generator and discriminator network were required, though these changes are very natural. The 3D-VAE-IWGAN image encoder mimics the simple 5 layer convolutional neural network encoder architecture described in . For the 3D-VAE-IWGAN voxel encoder, an architecture identical to that of the discriminator network described above for 3D generation is used, except that the final layer outputs a 400 dimensional vector of means and variances. Here one can more easily understand why it is necessary for the encoder and the discriminator to learn at the same pace. They receive very similar inputs from which we require that they learn practically the same features and at the same rate, so as to steadily push the generator in the right direction.
4.1 3D Generation
We first examined if our 3D-IWGAN system was a able to generate objects from a distribution consisting of only chairs, but set in 12 different orientations from the ModelNet10 dataset. As shown in Figure 4, our method produces objects of high quality from a variety of viewpoints. Figure 6 demonstrates that our methods is also able to interpolate between two chairs. As the latent variables shift from the first object to the last, there is not a sudden change, but a slow transition in which each segment can easily be identified as a chair. We next experimented with generating from the entire ModelNet10 dataset, with 10 object classes in 12 orientations simultaneously. This was also a markedly successful endeavor, as the the objects our trained models produced are successfully varied and of high quality. This can be verified in Figure 5. It was possible to track the quality of the objects using the discriminator’s loss, which can be viewed in Figure 7. This allowed the models convergence to be easily tracked and provided a clear signal for when to halt the learning. For the single object class test, the model trained for 4000 epochs, and for the multi-class test the model was trained for 1100 epochs.
The models trained here converged easily, with no tuning, and far more stably than with the original GAN framework. An example of this stable convergence can be viewed in Figure 7. We can see that the 3D-IWGAN system is a powerful generative model, consistently producing high quality objects, from distinctly difficult distributions.
To gain some understanding of how our 3D-IWGAN models deal with the transitions from one distinct object class and orientation to another, observe the bottom row of Figure 6. The transitions between the objects are all recognizable and clean. This is a remarkable change from training the original GAN system, which is well-known to induce large unusable areas between properly generated objects which produce unrealistic shapes . Our model’s more attractive transitions between objects demonstrate that it has learned a better organization of observed objects’ positions along the latent space, so as to produce as few of these unrealistic areas as possible. Our improved training scheme makes the unavoidable unrealistic interpolations represent as small a portion of the latent space as possible.
4.2 3D Object Reconstruction from Single Images
3D-IWGAN was applied to the task of recovering the 3D shape of objects found in images. This was a task also tackled by the 3DGAN system , and so provides fair and quantitative basis for comparison of the two methods. The results of our trained 3D-VAE-IWGAN system evaluated on the IKEA dataset can be observed in Table 1. Our system consistently outperforms the original 3DGAN system and several other previous approaches, with a mean average precision of 61.7 across all classes. Our jointly trained model also demonstrates improvement over those from all previous work, on average, and falls only slightly short of our new separately trained models. We note that a joint model faces a harder problem, intrinsically here, as the independent models are provided access to known category labels, while joint models are competitive in the fully unsupervised regime. These results demonstrate the large generative improvements which result from our new 3D-IWGAN system and the success of our new 3D-VAE-IWGAN system.
|T-L Network ||56.3||30.2||32.9||25.8||71.7||23.3||40.0|
|3D-VAE-GAN  jointly trained||49.1||31.9||42.6||34.8||79.8||33.1||41.2|
|3D-VAE-GAN  separately trained||63.2||46.3||47.2||40.7||78.8||42.3||53.1|
|3D-VAE-IWGAN (ours) jointly trained||65.7||44.2||49.3||50.6||68.0||52.2||55.0|
|3D-VAE-IWGAN (ours) separately trained||77.7||51.8||56.2||49.8||82.0||52.6||61.7|
4.3 Reconstruction of 3D Objects from Single Perspective Depth Scans
A voxel encoded variation of our 3D-VAE-IWGAN system was applied to the task of recovering the complete 3D shape of objects from the output of a single perspective scan from a depth sensor. Two models were produced; the first trained on only chair objects and the second on all the objects in the ModelNet10 dataset. It was apparent that both tests were quite successful and examples of recovered objects from the first and second tests set can be viewed in the second rows of Figure 8, and Figure 9.
To demonstrate that the models produced here are effective on actual Kinect scans, a set of real scans was evaluated as well. A small sample of Kinect depths maps, found in the A Large Dataset of Object Scans dataset  was converted into voxel representations, and passed though the trained GAN model for chair reconstruction. Two examples of this pipeline and the objects reconstructed through it can be viewed in Figure 10. This picture shows that our generative model accurately reconstructs the original chairs. The success in this task both on the synthetic models and true Kinect scan illustrates that the generative power of our 3D-VAE-IWGAN system makes it is easily applicable to tasks involving reproducing 3D shapes, from a variety of input types.
In this paper we outline a new system, 3D-IWGAN, which is successful in 3D generation from complex distributions involving multiple distinct classes; spaces which we highlight other 3D generative systems involving GANs cannot generalize to. We show the models produced by this system learn distributions involving detailed, multiple object classes in multiple orientations, and highlight that these models are easy to train, require no special tuning, and have a quantitative convergence. In addition, we describe how to combine this new system with Variational Auto-encoders to produce a 3D generative system, 3D-VAE-IWGAN, the output of which can be conditioned on some known data. We then demonstrate this system’s generative power by recovering 3D objects from images, to achieve state of the art performance on the IKEA dataset. Finally, we again demonstrate our system’s generative power and also its ability to be conditioned on complex data by successfully applying it to recovering objects’ 3D volume and shape from single perspective depth maps.
- Rezende et al.  D. J. Rezende, S. M. A. Eslami, S. Mohamed, P. Battaglia, M. Jaderberg, and N. Heess. Unsupervised learning of 3d structure from images. CoRR, abs/1607.00662, 2016. URL http://arxiv.org/abs/1607.00662.
- Sharma et al.  A. Sharma, O. Grau, and M. Fritz. Vconv-dae: Deep volumetric shape learning without object labels. CoRR, abs/1604.03755, 2016. URL http://arxiv.org/abs/1604.03755.
- Wu et al.  J. Wu, C. Zhang, T. Xue, B. Freeman, and J. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems, pages 82–90, 2016.
- Goodfellow et al.  I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014.
- Kingma and Welling  D. P. Kingma and M. Welling. Auto-Encoding Variational Bayes. ArXiv e-prints, Dec. 2013.
- Roberts  L. G. Roberts. Machine perception of three-dimensional solids. PhD thesis, 1963.
- Marr  D. Marr. Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. Henry Holt and Co., Inc., New York, NY, USA, 1982. ISBN 0716715678.
- Sung et al.  M. Sung, V. Kim, R. Angst, and L. Guibas. Data-driven structural priors for shape completion. ACM Transactions on Graphics, 2015.
- Sundar et al.  H. Sundar, D. Silver, N. Gagvani, and S. Dickinson. Skeleton based shape matching and retrieval. In Proceedings of the Shape Modeling International 2003, SMI ’03, pages 130–, Washington, DC, USA, 2003. IEEE Computer Society. ISBN 0-7695-1909-1. URL http://dl.acm.org/citation.cfm?id=829510.830339.
- Li et al.  C. Li, M. Z. Zia, Q.-H. Tran, X. Yu, G. Hager, and M. Chandraker. Deep supervision with shape concepts for occlusion-aware 3d object parsing. In Proceedings of the Conference on Computer Vision and Pattern Recognition, 2017.
- Kar et al.  A. Kar, S. Tulsiani, J. Carreira, and J. Malik. Category-specific object reconstruction from a single image. In Proceedings of the Conference on Computer Vision and Pattern Recognition, 2015.
- Dame et al.  A. Dame, V. A. Prisacariu, C. Y. Ren, and I. Reid. Dense reconstruction using 3d object shape priors. In Proceedings of the Conference on Computer Vision and Pattern Recognition, 2013.
- Dosovitskiy et al.  A. Dosovitskiy, J. T. Springenberg, and T. Brox. Learning to generate chairs with convolutional neural networks. In Proceedings of the Conference on Computer Vision and Pattern Recognition, 2015.
- Choy et al.  C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision, 2016.
- Wu et al.  Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shape modeling. In Conference on Computer Vision and Pattern Recognition, 2015.
- Denton et al.  E. L. Denton, S. Chintala, and R. Fergus. Deep generative image models using a laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems, 2015.
- Girdhar et al.  R. Girdhar, D. F. Fouhey, M. Rodriguez, and A. Gupta. Learning a predictable and generative vector representation for objects. CoRR, abs/1603.08637, 2016. URL http://arxiv.org/abs/1603.08637.
- Liu et al.  J. Liu, F. Yu, and T. Funkhouser. Interactive 3D Modeling with a Generative Adversarial Network. ArXiv e-prints, June 2017.
- Choi et al.  W. Choi, Y. W. Chao, C. Pantofaru, and S. Savarese. Understanding indoor scenes using 3d geometric phrases. In Proceedings of the Conference on Computer Vision and Pattern Recognition, 2013.
- Izadinia et al.  H. Izadinia, Q. Shan, and S. M. Seitz. Im2cad. In CVPR, 2017.
- Eigen and Fergus  D. Eigen and R. Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. CoRR, abs/1411.4734, 2014. URL http://arxiv.org/abs/1411.4734.
- Li et al.  J. Li, D. Meger, and G. Dudek. Learning to Generalize 3D Spatial Relationships. In Proceedings of the 2016 IEEE International Conference on Robotics and Automation, Stockholm, Sweden, May 2016.
- Li et al.  J. Li, D. Meger, and G. Dudek. Context-coherent scenes of objects for camera pose estimation. In Proceedings of the International Conference on Intelligent Robots and Systems, 2017.
- Arjovsky et al.  M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN. ArXiv e-prints, Jan. 2017.
- Gulrajani et al.  I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Courville. Improved training of wasserstein gans. CoRR, abs/1704.00028, 2017. URL http://arxiv.org/abs/1704.00028.
- Kingma and Ba  D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015.
- Larsen et al.  A. B. L. Larsen, S. K. Sønderby, and O. Winther. Autoencoding beyond pixels using a learned similarity metric. CoRR, abs/1512.09300, 2015. URL http://arxiv.org/abs/1512.09300.
- Lim et al.  J. J. Lim, H. Pirsiavash, and A. Torralba. Parsing IKEA Objects: Fine Pose Estimation. ICCV, 2013.
- Wu et al.  Z. Wu, S. Song, A. Khosla, X. Tang, and J. Xiao. 3d shapenets for 2.5d object recognition and next-best-view prediction. CoRR, abs/1406.5670, 2014. URL http://arxiv.org/abs/1406.5670.
- Chang et al.  A. X. Chang, T. A. Funkhouser, L. J. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, J. Xiao, L. Yi, and F. Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. URL http://arxiv.org/abs/1512.03012.
- Xiao et al.  J. Xiao, K. A. Ehinger, J. Hays, A. Torralba, and A. Oliva. Sun database: Exploring a large collection of scene categories. IJCV, 119(1):3–22, Aug. 2016. ISSN 0920-5691. doi:10.1007/s11263-014-0748-y. URL http://dx.doi.org/10.1007/s11263-014-0748-y.
- Cimpoi et al.  M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition, 2014.
- Choi et al.  S. Choi, Q.-Y. Zhou, S. Miller, and V. Koltun. A large dataset of object scans. arXiv:1602.02481, 2016.
- Radford et al.  A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. URL http://arxiv.org/abs/1511.06434.