Computer Vision

We are tackling fundamental problems in computer vision and graphics to achieve a deep understanding of visual information in the world.
Our main challenge in the computer vision domain is divided into two categories: generative and discriminative. Image generation tasks aim to create (or convert) an image reflecting the semantics of the given condition. The representative tasks are image-to-image translation, image colorization, image in-painting and video generation. The discriminative problem tries to compress an image into low-dimensional representation while preserving its essential information. It contains visual representation learning, label propagation, segmentation, and video understanding.

Deep Image Colorization
Deep image colorization is one of the deep generative tasks, which aims to generate a realistic color image given a sketch or grayscale image.
Due to its significant potentials for practical use in the content creation industries, deep image colorization has attracted a great deal of attention in the field of graphics and computer vision.
DAVIAN Lab. has recently proposed novel methods in various aspects, such as few-shot colorization via a memory network and exemplar-based sketch colorization using dense spatial correspondence.
We further attempt to alleviate the color-bleeding artifacts by enhancing edge-aware information via user-interactive edge hint and edge-refinement network.
Our ultimate goal in this task is to establish a human-ai collaborative colorization system for the integrated content creation.
Fashion items virtual fitting
Image-based virtual try-on is to synthesize a photo-realistic new image by overlaying a cloth image onto the corresponding region of a clothed person. Previous researches related to virtual try-on has a limitation to generate high-resolution images large enough to be used in real life. DAVIAN Lab. attempts to solve this problem by applying high-resolution conditional GANs.
Semantic segmentation
Semantic segmentation, a fundamental task in the field of computer vision, is one of the crucial research topics that have an impact on the automotive and medical industry. DAVIAN Lab. is conducting research in aspect of domain generalization in order to make a robust semantic segmentation algorithm that is practically necessary for the industry. Additionally, we are also interested in research on how to exploit the strucutral priors existing in images to improve segmentation performance.