Eugene Siow commited on
Commit
3c60ebd
1 Parent(s): 756f356

Add update to dataset curator and source information.

Browse files
Files changed (1) hide show
  1. README.md +28 -6
README.md CHANGED
@@ -52,7 +52,6 @@ task_ids:
52
  - **Repository**: https://huggingface.co/datasets/eugenesiow/Div2k
53
  - **Paper**: http://www.vision.ee.ethz.ch/~timofter/publications/Agustsson-CVPRW-2017.pdf
54
  - **Leaderboard**: https://github.com/eugenesiow/super-image#scale-x2
55
- - **Point of Contact**: radu.timofte@vision.ee.ethz.ch
56
 
57
  ### Dataset Summary
58
 
@@ -139,21 +138,44 @@ The data fields are the same among all splits.
139
 
140
  #### Initial Data Collection and Normalization
141
 
142
- [More Information Needed]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
143
 
144
  #### Who are the source language producers?
145
 
146
- [More Information Needed]
 
147
 
148
  ### Annotations
149
 
150
  #### Annotation process
151
 
152
- [More Information Needed]
153
 
154
  #### Who are the annotators?
155
 
156
- [More Information Needed]
157
 
158
  ### Personal and Sensitive Information
159
 
@@ -179,7 +201,7 @@ immediately.
179
 
180
  ### Dataset Curators
181
 
182
- [More Information Needed]
183
 
184
  ### Licensing Information
185
 
 
52
  - **Repository**: https://huggingface.co/datasets/eugenesiow/Div2k
53
  - **Paper**: http://www.vision.ee.ethz.ch/~timofter/publications/Agustsson-CVPRW-2017.pdf
54
  - **Leaderboard**: https://github.com/eugenesiow/super-image#scale-x2
 
55
 
56
  ### Dataset Summary
57
 
 
138
 
139
  #### Initial Data Collection and Normalization
140
 
141
+ **Resolution and quality**: All the images are 2K resolution, that is they have 2K pixels on at least one of
142
+ the axes (vertical or horizontal). All the images were processed using the same tools. For simplicity, since the most
143
+ common magnification factors in the recent SR literature are of ×2, ×3 and ×4 we cropped the images to multiple of
144
+ 12 pixels on both axes. Most of the crawled images were originally above 20M pixels.
145
+ The images are of high quality both aesthetically and in the terms of small amounts of noise and other corruptions
146
+ (like blur and color shifts).
147
+
148
+ **Diversity**: The authors collected images from dozens of sites. A preference was made for sites with freely
149
+ shared high quality photography (such as https://www.pexels.com/ ). Note that we did not use images from Flickr,
150
+ Instagram, or other legally binding or copyright restricted images. We only seldom used keywords to assure the diversity
151
+ for our dataset. DIV2K covers a large diversity of contents, ranging from people, handmade objects and environments
152
+ (cities, villages), to flora and fauna, and natural sceneries including underwater and dim light conditions.
153
+
154
+ **Partitions**: After collecting the DIV2K 1000 images the authors computed image entropy, bit per pixel (bpp) PNG
155
+ compression rates and CORNIA scores (see Section 7.6) and applied bicubic downscaling ×3 and then upscaling ×3 with
156
+ bicubic interpolation (imresize Matlab function), ANR [47] and A+ [48] methods and default settings.
157
+
158
+ The authors randomly generated partitions of 800 train, 100 validation and 100 test images until they achieved a good
159
+ balance firstly in visual contents and then on the average entropy, average bpp, average number of pixels per
160
+ image (ppi), average CORNIA quality scores and also in the relative differences between the average PSNR scores of
161
+ bicubic, ANR and A+ methods.
162
+
163
+ Only the 800 train and 100 validation images are included in this dataset.
164
 
165
  #### Who are the source language producers?
166
 
167
+ The authors manually crawled 1000 color RGB images from Internet paying special attention to the image quality,
168
+ to the diversity of sources (sites and cameras), to the image contents and to the copyrights.
169
 
170
  ### Annotations
171
 
172
  #### Annotation process
173
 
174
+ No annotations.
175
 
176
  #### Who are the annotators?
177
 
178
+ No annotators.
179
 
180
  ### Personal and Sensitive Information
181
 
 
201
 
202
  ### Dataset Curators
203
 
204
+ - **Original Author**: [Radu Timofte](http://people.ee.ethz.ch/~timofter/)
205
 
206
  ### Licensing Information
207