Auto-colorization Exploiting Annotated Dataset

Colorization is a very challenging task for computers which requires very high performance of segmentation, object recognition, color understanding, etc., and even for humans, it’s very demanding and arduous work to fully accomplish. To achieve this task, there traditionally has been three approaches: example based model, scribble based model, and data-driven based model which introduced relatively recently, but none of those algorithms are humanintervention free, nor generically work for any arbitrary images. In this paper, I introduce a data-driven based autocolorization algorithm which is applicable for an arbitrary query image without extra human labors. I set a hypothesis that if you have many similar images to a query image, a query image is highly likely to have the same objects as in the similar images. To prove this, I exploited the biggest annotated dataset in the world, LabelMe [12], and built a statistical model estimating likelihood of objects in the query image. Once I have a probabilistic map, I simply transfer the colors to each map by using filling algorithm introduced by Levin et al. [4]