NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study

By Eirikur Agustsson, Radu Timofte. 2017

4,043 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition, Media Technology, Evaluation & Benchmarks

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

In Media Technology

In Evaluation & Benchmarks

Citations in the corpus

High-Resolution Image Synthesis with Latent Diffusion Models (2021, methodology): “Images are from the DIV2K [1] validation set, evaluated at 512(2) px.”
Contrastive Multiview Coding (2019, neutral): “2k) from DIV2K [2] dataset.”
Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (2021, methodology): “Speciﬁcally, the four degradation types for DIV2K4D including 1) type I: the commonly-used bicubic degradation; 2) type II: anisotropic Gaussian blur with nearest downsampling by a scale fac-( tor of 4; 3) type III: anisotropic Gaussian blur with nearest downsampling by a scale factor of 2 and subse”
The Unreasonable Effectiveness of Deep Features as a Perceptual Metric (2018, methodology): “Superresolution We evaluate results from the 2017 NTIRE workshop [2].”
Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains (2020, methodology): “The Natural images are generated by taking center crops of randomly sampled images from the Div2K dataset [1].”
What makes for good views for contrastive learning (2020, background): “We increase d from 64 to 384, and sample patches from inside high resolution images in the DIV2K dataset [1].”
Fast Lossless Neural Compression with Integer-Only Discrete Flows (2022, methodology): “We evaluate the models’ compression performance on the high-resolution image dataset CLIC 4 ( ∼ 10 6 pixels per image) (Agustsson & Timofte, 2017) and compare the performance with non-neural compression algorithms.”
High-Fidelity Generative Image Compression (2020, methodology): “We use a patch size of 256 pixels, which yields 28 650 patches for CLIC2020 and 6 573 patches for DIV2K.”
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks (2018, methodology): “For training, we mainly use the DIV2K dataset [40], which is a high-quality (2K resolution) dataset for image restoration tasks.”
Palette: Image-to-Image Diffusion Models (2021, neutral): “Note that this is unlike most prior work [Dong et al. 2015; Liu et al. 2018b], which mainly use small datasets such as DIV2K [Agustsson and Timofte 2017] and BSD500 [Martin et al. 2001] for training and evaluation.”
Deep Interactive Thin Object Selection (2021, methodology): “Speciﬁcally, for training and validation set, we use HRSOD [52] and DIV2K [3] as high-resolution backgrounds, and COCO [26] as low-resolution backgrounds.”
Image Super-Resolution via Progressive Cascading Residual Network (2018, methodology): “And to construct intermediary scale images for progressive training, we use DIV2K {×2,×4} dataset [1].”
StegaStamp: Invisible Hyperlinks in Physical Photographs (2019, methodology): “The network is trained using a dataset of randomly transformed StegaStamps embedded into high resolution images sampled from DIV2K [1].”
Neural Nearest Neighbors Networks (2018, methodology): “For training, we use the 400 images of the BSDS training and test splits, 800 images of the DIV2K training set [1], and a training split of 3793 images from the Waterloo database [29].”
Efficient Long-Range Attention Network for Image Super-resolution (2022, methodology): “An enhanced residual block was proposed in [29] to train the deep model without batch normalization, and the developed EDSR network won the first prize of the NTIRE2017 challenge [1].”
Towards Flexible Blind JPEG Artifacts Removal (2021, methodology): “Following [13], we employ DIV2K [1] and Flickr2K [38] as our training data.”