LR-GAN: Layered Recursive Generative Adversarial Networks for Image Generation

By Jianwei Yang, Anitha Kannan, Dhruv Batra, Devi Parikh. 2016

247 citations recorded by the source corpus.

We present LR-GAN: an adversarial image generation model which takes scene structure and context into account. Unlike previous generative adversarial networks (GANs), the proposed GAN learns to generate image background and foregrounds separately and recursively, and stitch the foregrounds on the background in a contextually relevant manner to produce a complete natural image. For each foreground, the model learns to generate its appearance, shape and pose. The whole model is unsupervised, and is trained in an end-to-end manner with gradient descent methods. The experiments demonstrate that LR-GAN can generate more natural images with objects that are more human recognizable than DCGAN.

Subfields: Computer Graphics and Computer-Aided Design, Computer Vision and Pattern Recognition, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Graphics and Computer-Aided Design

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

Spectral Normalization for Generative Adversarial Networks (2018, neutral): “(2017)), ‡ (Yang et al., 2017), ∗ (Warde-Farley & Bengio, 2017), †† (Gulrajani et al.”
Classification Accuracy Score for Conditional Generative Models (2019, methodology): “To that end, we use a general framework (first posed in [12] and further studied in [13]) in which we use conditional generative models to perform approximate inference and measure the quality of that inference.”
Population Based Training of Neural Networks (2017, neutral): “45 , similar to or better than Inception scores for DCGAN-like architectures reported in prior work (Rosca et al., 2017; Gulrajani et al., 2017; Yang et al., 2017).”
Tell, Draw, and Repeat: Generating and Modifying Images Based on Continual Linguistic Instruction (2018, methodology): “[13] perform unsupervised image generation in recursive steps, first generating a background, subsequently conditioning on it to generate the foreground and the mask, and finally using an affine transformation to combine the foreground and background.”
Pros and Cons of GAN Evaluation Measures (2018, methodology): “[95] proposed two metrics based on the intuition that a sufficient, but unnecessary, condition for closeness of generated data distribution Pg(x) and the real data distribution Pr(x) is closeness of Pg(x|y) and Pr(x|y), i.”
DA-GAN: Instance-Level Image Translation by Deep Attention Generative Adversarial Networks (2018, neutral): “Compared with existing works which uses foreground masks LRGAN [40] or normal maps SSGAN [38] to retain the identity in a specific task, the proposed DA-GAN can automatically and adaptively learn task-driven identity representations by attention mechanism without human-involvement.”
Compositional GAN: Learning Image-Conditional Binary Composition (2018, methodology): “Generative adversarial networks (GANs) have been used inawidevarietyofsettingsincludingimagegeneration(Den-ton et al. 2015; Yang et al. 2017; Karras et al. 2017) and representation learning (Radford et al. 2016; Salimans et al. 2016; Liu and Tuzel 2016; Chen et al. 2016).”
Bidirectional Conditional Generative Adversarial Networks (2017, methodology): “model. We show the accuracy and the F 1-score of these external models on real test datasets for reference as AExt real and F Ext real . We also calculate the adversarial accuracy (AA) as proposed in [33]. 2https://github.com/bstriner/keras-adversarial 3https://github.com/fchollet/keras/blob/master/”
Unsupervised Attention-guided Image to Image Translation (2018, background): “’s recursive approach [23] generates images by decoupling the generation of the foreground and background in a sequential manner; however, its extension to image-to-image translation is not straightforward as in that case we only care about modifying the foreground.”
MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks (2019, methodology): “More recently, a number of GAN-based methods break the process of high resolution image synthesis into smaller subtasks [36, 39, 38, 8, 9, 15, 34, 40].”
Unsupervised Object Segmentation by Redrawing (2019, background): “Most of those can either only be used in generation [56], or only handle very simple objects [4, 14, 18].”
On the Evaluation of Conditional GANs (2019, background): “Most of these metrics either focus on the separability between generated images and real images [21, 30, 50, 16], compute the distance between distributions [10, 12, 2], assess sample quality and diversity from conditional or marginal distributions [33, 11, 55], measure the similarity between genera”
MixNMatch: Multifactor Disentanglement and Encoding for Conditional Image Generation (2019, unknown): “sketch images. Inception Score FID Birds Dogs Cars Birds Dogs Cars Simple-GAN 31.85 0.17 6.75 0.07 20.92 0.14 16.69 261.85 33.35 InfoGAN [4] 47.32( 0.77 43.16 0.42 28.62 0.44 13.20 29.34 17.63 LR-GAN [45] 13.50 0.20 10.22 0.21 5.25 0.05 34.91 54.91 88.80 StackGANv2 [48] 43.47 0.74 37.29 0.56 33.69 0”
Predicting Deeper into the Future of Semantic Segmentation (2017, unknown): “[45] use similar ideas to develop an iterative im-”