Deconvolution and Checkerboard Artifacts

By Augustus Odena, Vincent Dumoulin, Chris Olah, Christopher Olah. 2016

1,756 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

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

In Computer Vision and Pattern Recognition

Citations in the corpus

Making Convolutional Networks Shift-Invariant Again (2019, neutral): “Similar to the subsampling case, this leads to aliasing, in the form of grid artifacts (Odena et al., 2016).”
Large Scale GAN Training for High Fidelity Natural Image Synthesis (2018, background): “Most previous failures involve local artifacts (Odena et al., 2016), images consisting of texture blobs instead of objects (Salimans et al.”
Regionwise Generative Adversarial Image Inpainting for Large Missing Areas (2019, background): “lly reasonable contents ﬁlled in the missing regions. However, it is very common to produce unwanted artifacts in unstable generative models, which could cause visually unrealistic results. Some work [53], [54] tried to analyze the reason behind, but still cannot totally overcome this problem. Image”
EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning (2019, methodology): “[40] to be an effective tool to combat “checkerboard” artifacts caused by transpose convolution layers [37].”
Conditional Image Synthesis with Auxiliary Classifier GANs (2016, background): “Broadly speaking, the architecture of the generator G is a series of ‘deconvolution’ layers that transform the noise z and class c into an image (Odena et al., 2016).”
Adversarially Learned Inference (2016, neutral): “nvolution, i.e., the gradient of the convolution with respect to its inputs. For more details about transposed convolutions and related operations, see Dumoulin & Visin (2016); Shi et al. (2016); Odena et al. (2016). 4.1 SAMPLES AND RECONSTRUCTIONS For each dataset, samples are presented (Figure”
Assessing Generative Models via Precision and Recall (2018, background): “compared based on statistical regularities in the images rather than random artifacts resulting from the generative process [17, 19].”
Photographic Image Synthesis with Cascaded Refinement Networks (2017, methodology): “Note that we do not use upconvolutions because upconvolutions tend to introduce characteristic artifacts [30].”
Learning-based Video Motion Magnification (2018, methodology): “We downsample with the strided convolution [20], and we use nearest-neighbor upsampling followed by a convolution layer to avoid checkerboard artifacts [21].”
Latent Space Physics: Towards Learning the Temporal Evolution of Fluid Flow (2018, background): “We assume the input is padded, and hence for s = 1 the output size matches the input, while choosing a stride s > 1 results in a downsampled output [ODO16].”
MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis (2019, methodology): “Checkerboard artifacts As noticed in Odena et al. (2016), deconvolutional generators are susceptible to generating "checkerboard" patterns if the kernel-size and stride of the transposed convolutional Figure 1: MelGAN model architecture.”
Parallel Wavegan: A Fast Waveform Generation Model Based on Generative Adversarial Networks with Multi-Resolution Spectrogram (2019, background): “D convolutions so that the time-resolution of auxiliary features matches the sampling rate of the speech waveform [11,28].”
Visual Foresight: Model-Based Deep Reinforcement Learning for Vision-Based Robotic Control (2018, background): “Standard subsampling and upsampling between convolutions is known to produce artifacts for dense image generation tasks [51], [52].”
Adversarial Audio Synthesis (2018, methodology): “We compare this to the performance of WaveGAN with phase shuffle (n ∈ {2, 4}) and a variant of WaveGAN which uses nearest-neighbor upsampling rather than transposed convolution (Odena et al., 2016).”
Unsupervised Pixel-Level Domain Adaptation with Generative Adversarial Networks (2016, neutral): “The discriminator D is a convolutional neural network where the number of layers depends on the image resolution: the first layer is a stride 1x1 convolution (motivated by [33]), which is followed by repeatedly stacking stride 2x2 convolutions until we reduce the resolution to less or equal to 4x4.”