Kwang Hee Won

Natural image matting based on neighbor embedding

In this paper, an automatic technique for natural image matting is proposed. We use visual characteristics of the background and foreground regions for matting. Through the Locally Linear Embedding (LLE) of high-dimensional features, we estimate foreground and background color of unknown pixels. We use gradient information and a hierarchical model to enhance the proposed matting technique. Instead of a user interaction for tri-map generation, we propose an automatic technique that obtains a tri-map using a multi-view camera. A reliability map for the depth of a scene facilitates the generation of a tri-map. It is proven that feature sets obtained from training images can be applied to similar images or video frames.

Geometric feature selection for vehicle pose estimation on dynamic road scenes

This paper suggests classification and selection of features based on geometric properties to accomplish more accurate visual odometry. We apply billboard sweep stereo algorithm and the homogeneous disparity representation to obtain several categories of geometric labels. And we combine the classes to estimate more accurate vehicle pose. In the experiments, the distant features are useful for estimating vehicle orientation and the features on the ground plane are proper for estimating vehicle position especially in crowded road scenes.

hSGM: hierarchical pyramid based stereo matching algorithm

In this paper, we propose a variant of Semi-Global Matching, hSGM which is a hierarchical pyramid based dense stereo matching algorithm. Our method aggregates the matching costs from the coarse to fine scale in multiple directions to determine the optimal disparity for each pixel. It has several advantages over the original SGM: a low space complexity and efficient implementation on GPU. We show several experimental results to demonstrate our method is efficient and obtains a good quality of disparity maps.

Pedestrian detection using labeled depth data

This paper presents pedestrian detection algorithm on labeled depth data which is obtained from road scenes. Our approach computes feature responses for head and legs of human body using depth and label data. And then, it detects pedestrians by removing edges and partitioning a bipartite graph of head and leg response blobs using prior knowledge about human body. In the experiments, the proposed algorithm produces better result compared to the method which uses histogram of gradient feature and the ground plane for road scenes.

Adaptive and scalable color correction for stereoscopic 3D displays

This paper presents a color correction algorithm on a stereo video using a locality-based color correspondence search. The adaptive and iterative refinement process ensures the robust color correction. It is real-time on a single-core CPU.

Rotation estimation for visual odometry using 3D vector correspondence

Visual odometry is important to obtain the vehicle motion as well as the relative motion of the surroundings such as moving/stationary objects to the vehicle. This paper presents a vehicle motion estimation approach which uses 3D vector correspondences. We design an objective function using 3D vector registration error and optimize it. The proposed method estimates more accurate rotation matrix compared to the previous method.

Adaptive local color correction for stereoscopic three-dimensional displays

Abstract. A color correction algorithm for stereoscopic video that uses a locality-based color correspondence search is presented. The algorithm overcomes the common failure problems that affect conventional local methods by introducing two techniques. First, an adaptive locality selection technique is introduced to determine a local image window, which is used to define an affine color correction model. The color correction model is then refined iteratively using locality-based color correspondence search between stereo video frames. Second, the color correspondence samples from the current video frame are used for the temporally stable color correction of the next frame. The experimental results show that the proposed algorithm outperforms the conventional local methods in terms of accuracy and time.

Localized earth mover's distance for robust histogram comparison

The Earth Movers Distance (EMD) is a useful cross-bin distance metric for comparing two histograms. The EMD is based on the minimal cost that must be paid to transform one histogram into the other. But outlier noise in the histogram causes the EMD to be greatly exaggerated. In this paper, we propose the localized Earth Movers Distance (LEMD). The LEMD separates noises from meaningful transportation of data by specifying local relations among bins, and gives a predefined penalty to those noises, according to the applications. An extended version of the tree-based transportation simplex algorithm is proposed for LEMD. The localized property of LEMD is formulated similarly to the original EMD with the thresholded ground distance, such as EMD-hat [7] and FastEMD [8]. However, we show that LEMD is more stable than EMD-hat for noise-added or shape-deformed data, and is faster than FastEMD that is the state of the art among EMD variants.

Automatic Stroke Extraction and Stroke Ordering Based on TrueType Font

요 약 이 논문에서는 트루타입 폰트의 글자 외곽선 데이터를 이용하여 자동으로 한 자의 획을 분리하고 획 순서를 정하는 방법을 제안한다. 트루타입 폰트에는 글자의 외곽선 정보가 벡터 형식으로 저장되어 있으며, 이러한 벡터들은 일 정한 규칙으로 배열되어 있다. 이와 같은 벡터들의 배치를 이용하여 한자의 획이 될 수 있는 벡터들의 집합을 조합하여 독립적인 획을 분리해 내고, 글 자를 획 별로 분리하여 본래 트루타입 폰트의 저장 형식과 동일한 파일 형식 으로 저장한다. 또한 분리된 모든 획에 대하여, 획 이름을 정의하고, 정의된 획들 간의 위치와 상관관계를 이용하여, 획 사이의 우선순위를 결정하여 획 순서를 부여한다. 이 작업들은 사람의 작업 없이 순수하게 자동으로 이루어 지므로, 시간과 노력을 최소화 할 수 있다. 게다가, 획 별로 분리되고 순서대 로 정리된 한자들은 트루타입 폰트에 저장되어 있는 모양과 특성을 그대로 가지고 있으므로, 단순히 폰트 자체로써 사용할 수도 있을 뿐만 아니라, 한자 학습 컨텐츠로도 이용이 가능하며, 각종 애니메이션 효과 등 다양한 분야에 서 융통성 있게 활용될 수 있다.