Label to Region by Bi-Layer Sparsity Priors

这是我目前找到的最早关于region tagging的文章，这里叫Label to Region。

The basic philosophy is that an image or semantic region can be sparsely reconstructed via the image patches belonging to the images with common image labels。

 (1) we cannot directly use visual similarity between the candidate region and atomic patch to select patches for the reconstruction purpose, since an atomic patch is only part of a region and their similarity cannot convey the inclusion relations; and (2) an intuitive way to improve the accuracy of cross-image region-to-region correspondence is to enforce the usage of atomic patches from few images for this reconstruction, from which those selected atomic patches from one image may have high possibility to form a semantic region

The proposed label-to-region assignment process has the following characteristics: 1) the bi-layer sparse coding aims to enforce the usage of merged patches within an image to reconstruct the reference image or semantic region, which ensures the reliability of label propagation; 2) the process does not require exact image object/concept parsing, which is still far from satisfactory on real world images; and 3) no generative model for each label/concept is learnt, and thus it is scalable to applications with large label set. 

Two images annotated with common labels are likely to contain some similar patches. However it is generally difficult to directly derive those semantically similar patch pairs between two images

Thus instead we use a group of atomic patches to reconstruct an image or semantic region, and then harness the reconstruction coefficients for propagating the image labels to those localized image patches。（很多patches不能完整地表达一个label ，或者说能同时表达好几个label）

Two images annotated with common labels are likely to contain some similar patches. However it is generally difficult to directly derive those semantically similar patch pairs between two images。（如果不仅仅考虑图片数量少，而且考虑图片的代表性）