PatchMatch: a randomized correspondence algorithm for structural image editing

By Connelly Barnes, Eli Shechtman, Adam Finkelstein, Dan B. Goldman. 2009

3,316 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

Non-local Neural Networks (2017, background): “Non-local matching is also the essence of successful texture synthesis [12], super-resolution [16], and inpainting [1] algorithms.”
Context Encoders: Feature Learning by Inpainting (2016, methodology): “Figure 5: Comparison with Content-Aware Fill (Photoshop feature based on [2]) on held-out images.”
Instance-Sensitive Fully Convolutional Networks (2016, methodology): “This is made possible by exploiting local coherence [9] of natural images for generating per-window pixel-wise predictions.”
Large Scale Image Completion via Co-Modulated Generative Adversarial Networks (2021, methodology): “Traditional methods (Ballester et al., 2001; Barnes et al., 2009; Darabi et al., 2012; Efros & Freeman, 2001; Efros & Leung, 1999) utilize only low-level features and fail to generate semantically consistent contents.”
Regionwise Generative Adversarial Image Inpainting for Large Missing Areas (2019, methodology): “Patch-based methods [20], [21] ﬁll in missing regions with patches from a collection of source images that maximize the patch similarity.”
Semantic Image Synthesis With Spatially-Adaptive Normalization (2019, methodology): “ditional image synthesis refers to the task of generating photorealistic images conditioning on some input data. Earlier methods compute the output image by stitching pieces from a database of images [3,13]. Recent methods directly learn the mapping using neural networks [4,7,20,39,40,45,46,47]. The”
EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning (2019, methodology): “PatchMatch [3] addressed this issue by using a fast nearest neighbor field algorithm.”
Free-Form Image Inpainting With Gated Convolution (2018, methodology): “Traditional approach PatchMatch [Barnes et al. 2009] copies similar patches into holes.”
Generative Visual Manipulation on the Natural Image Manifold (2016, methodology): “More advanced editing methods such as image warping [8,9] or structured image editing [10] intelligently reshuffle the pixels in an image following user’s edits.”
Semantic Image Inpainting with Deep Generative Models (2016, background): “PM [2] cannot be applied to this case due to insufficient available patches.”
Rewriting a Deep Generative Model (2020, background): “Common operations include color transfer [65,53], image deformation [67,75], object cloning [62,11], and patch-based image synthesis [20,5,29].”
MaskGIT: Masked Generative Image Transformer (2022, background): “Traditional patch-based methods [3] work well on texture regions, while deep learning based methods [14,38,48,50,54] have been demonstrated to synthesize images requiring better semantic coherence.”
SurfaceNet: An End-to-End 3D Neural Network for Multiview Stereopsis (2017, methodology): “The depth map fusion algorithms first recover depth maps [30] from view pairs by matching similarity patches [2, 18, 33] along epipolar line and then fuse the depth maps to obtain a 3D reconstruction of the object [27, 4, 9].”
3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction (2016, methodology): “We used a Patch-Match [48]-based oﬀ-the-shelf implementation [49] as the MVS method.”
Palette: Image-to-Image Diffusion Models (2021, methodology): “Baselines: ‡Photoshop’s Content-aware Fill, based on PatchMatch [Barnes et al., 2009], †[Yu et al.”
View Synthesis by Appearance Flow (2016, background): “The notion of appearance flow field is closely related to the nearest neighbor field (NNF) in PatchMatch [26], except that NNF is explicitly defined on a distance function between two patches, while our appearance flow field is the output of a CNN after end-to-end training for cross-view reconstruct”