Vision is the most complicated and high-performing sense that humans have. In fact, we are able to understand the present or the past and predict the future simply by taking a look of the surrounding objects without even touching or feeling them. Such visual capacity that is inherent in us enables us to continue a conversation by adjusting to the other person’s mood by looking at the facial expressions and gestures. It also enables a baseball player to hit a fast ball thrown by a pitcher. Unlike humans, however, it is extremely difficult to give this visual intelligence to robots, motor vehicles, and electronic appliances. Setting the realization of visual intelligence required for cognition, judgment, and execution as our main goal, we endeavor to make human life more convenient and safe.
Visual AnalyticsA lot of research has been done on object classification and detection, which are the basic element technologies in the traditional computer vision field. It is also widely applied in
fields such as vision inspection, video surveillance (CCTV), and autonomous driving. Although a large amount of data is required to train a deep learning model, there are fields of
applications where it is difficult to collect data, and there is even an annotation bottleneck in which an answer must be labeled to a large amount of data for supervised learning.
To overcome these data-related limitations and efficiently apply AI to application fields, we are conducting research such as Self-Supervised Representation Learning that learns
the characteristics of image data without labeling, Continual Learning to solve the phenomenon where earlier learned information is forgotten (catastrophic forgetting), and learn to
perform a new task, Active Learning that can learn quickly with only a small number of datasets by efficiently selecting informative data and repeatedly performing human labeling, and
Explainable AI that can visualize the judgment basis of deep learning classification models.
Visual UnderstandingVisual Understanding abstracts visual information with the goal of making it into data that AI or humans can understand. Through this, we aim to make AI thinking and decision-making
similar to that of human based on a comprehensive and higher understanding of the surrounding environment. Research on technologies such as Object Detection, Segmentation, Graph Networks, Transformer, and OCR are being carried out, which are utilized as Deep Document Understanding tasks that play a pivotal role in LG Group’s digital transformation.
Large Scale Vision ModelRecently, researches on Large-scale Scale Vision Model in the field of Vision AI are being actively conducted. These researches are exploring various possibilities that can be incorporated into our
daily lives. We are conducting research from Image and Language-Based Multi-Modal Representation models to Image and Text Generation models.
3D VisionResearches are being carried out to enable a vivid experience in virtual humans and virtual reality, aiming to create natural lip-sync and facial expressions of digital human. In addition, the
research is being conducted on AI models that perform dynamic modeling on various movements and generate natural movements based on them.
Meet our leader!
Seunghwan Kim, Head of Vision Lab
“The Vision Lab conducts a wide range of vision research from Image/Video Analytics, Understanding, Generation to Low-level Vision and 3D Reconstruction & Synthesis.
I highly suggest you to apply to LG AI Research and conduct your research using our millions of datasets as a stepping stone and create a new future with vision researchers who are eager to grow together with LG AI Research.”