From Patches to Pictures (PaQ-2-PiQ): Mapping the Perceptual Space of Picture Quality

To advance progress on this problem, we introduce the largest subjective picture quality database, containing about 40000 real-world distorted pictures and 120000 patches, on which we collected about 4M human judgments of picture quality

Zhenqiang Ying; Haoran Niu; Praful Gupta; Dhruv Mahajan; Deepti Ghadiyaram; Alan Bovik

2020

Scholarcy highlights

  • Often of questionable quality, have become ubiquitous. Several hundred billion photos are uploaded and shared annually on social media sites like Facebook, Instagram, and Tumblr
  • Some FR algorithms have achieved remarkable commercial success, they are limited by their requirement of pristine reference pictures
  • As is the common practice in the field of picture quality assessment, we report two metrics: Spearman Rank Correlation Coefficient and Linear Correlation Coefficient
  • From Table 5, the first thing to notice is the level of performance attained by popular shallow models, which have the same feature sets
  • We developed a model that produces local quality inferences, uses them to compute picture quality maps, and global image quality
  • We applied the following criteria when randomly cropping out patches: aspect ratio: patches have the same aspect ratios as the pictures they were drawn from. dimension: the linear dimensions of the patches are 40%, 30%, and 20% of the picture dimensions. location: every patch is entirely contained within the picture, but no patch overlaps the area of another patch cropped from the same image by more than 25%
  • We believe that the proposed new dataset and models have the potential to enable quality-based monitoring, ingestion, and control of billions of social-media pictures and videos

Need more features? Save interactive summary cards to your Scholarcy Library.