Network Ensemble (Fusion)

Network ensemble can be said as the process of making a single prediction by combining the predictions from multiple networks. The field of network ensemble has been studied from the time when machine learning technology began to develop, but research was focused mainly on simple classification, CNNs with low complexity, or weak predictors. Research on ensembles deep neural networks (DNNs) has been very slow. Thus a novel method of fusing multiple DNNs to improve performance (and reliability) without modifying nor retraining the networks is required.

Object Detection

The role of object detection is to localize (position, size) semantic objects having a certain class (e.g. humans, cars, animals, or etc) in digital images. Object detection is widely used in the field of computer vision because of its wide applications to video surveillance, image retrieval, human-computer interface, medical imaging, and so on.

Driver Assistant System (DAS)

Driver assistance system is an automated vehicle system to support safe driving by analyzing dynamic environment. The main applications are ACC (Adaptive Cruise Control), FCW (Forward Collision Warning), AHB (Adaptive High-Beam). In the viewpoint of vision technology, they commonly consist of 3D object detection/tracking, dense stereo matching, and visual odometry.

SLAM and Structure from Motion

SLAM is a technique used by robots and autonomous vehicles and is defined as the problem of building a model leading to a new map, or repetitively improving an existing map, while at the same time localizing the robot within that map.

Object Categorization based on Graph Structures

Object Categorization is a classical research topic which determines whether the given image contains the certain category of object. Because the objects are represented with different color, shape, structure in images, we need to understand the characteristics of each object class and describe them effectively. In order to address these variations, the graph structure could be the one of solutions because the graph structure has a strong capacity for representation, and it is robust to various deformations. In this research, we study about how generate the representative graph model for each category and recognize them.

IMU-camera based Navigation

Fusion of heterogeneous sensors is an important issue to handle practical issues in a real world. For the localization of mobile devices, the IMU-camera fusion which is inspired from the human system is very useful due to the complimentary properties.

Color Correction for Image Mosaic

Many computer vision tasks, such as image matching and image stitching, require color consistency. However, different cameras or settings could generate different colored images despite a static scene. To be similar for colors of multiple images, methods presented in this field change colors of input images. Here, we focus on handling color differences in the case of image stitching.