Guang Jiang

A new fast motion estimation algorithm based on the loop-epipolar constraint for multiview video coding

There lie geometric constraints between neighboring frames in multiview video sequences. The geometric constraints are valuable for reducing spatial and temporal redundancy in multiview video coding (MVC). In this paper, we propose a new fast motion estimation algorithm based on the loop-epipolar constraint which combines loop and epipolar constraints. A practical search technique is designed according to the characteristics of the loop-epipolar constraint. Experimental results show that the proposed algorithm is efficient for sequences under different multiview camera setups. Highlights? A fast motion estimation algorithm is proposed based on the loop-epipolar constraint. ? The loop-epipolar constraint combines loop and epipolar constraints. ? A practical search technique is designed according to the characteristics of the loop-epipolar constraint. ? The proposed algorithm is efficient for different sequences.

Mean shift tracking with graph cuts based image segmentation

In this paper, we propose a new tracking method that applies image segmentation based on graph cuts to mean shift tracking algorithm. Mean shift tracking algorithm is an iterative scheme based on comparing the color histogram of the original object in the current image frame and the color histogram of candidate regions in the next image frame. Graph cuts can be employed to efficiently solve the problem of image segmentation in computer vision. Before tracking the object in each image frame of a video sequence, the object can be separated from the uninterested background using graph cut. Thus, the mean shift algorithm can obtain a more precise object histogram model of each frame without the interference of the background-pixels. The tracking window is designated by users and the segmenting window is generated through the tracking window. The proposed method provides more reliable performance than general mean shift tracking algorithm.

An epipolar geometry guided image inpainting method

Image inpainting, which is an important topic in the fields of computer vision and image processing, has been widely investigated in many applications. But traditional inpainting methods usually fail to get a reliable effect because of the lacking of the information. In this paper, we present a new inpainting method with considering a reference image exist. Different from the single view inpainting, we can get the information of the damaged regions from the reference image. Epipolar geometry constraint between the two images is added to the exemplar-based image inpainting algorithm. Experimental results show that our method can repair both structures and textures of the missing regions with a reliable result.

Mean-Shift Tracking of Variable Kernel Based on Projective Geometry

The Mean-shift algorithm is very useful in object tracking for its many advantages, such as good performance in real-time tracking, nonparametric density model, etc. Although the scale of the mean-shift kernel is a crucial parameter, there exists presently still no clear mechanism in choosing or updating the scale when the kernel of changing size is tracked. In this paper, a new method is introduced using projective geometry to determine the kernel size of the object. After initialization of this algorithm, we obtain the geometric information, and decide the corresponding kernel size of the object wherever the object moves. The experimental results show that this algorithm works stably and it consumes less time than traditional algorithms. I. INTRODUCTION

Registration of point clouds from kinect with congruent spheres

In this paper, a novel method for solving the registration of point clouds from Kinect is proposed. The proposed method, which requires neither initializations nor iterations, is totally different from the ICP algorithm. The basic idea of this method is to use the congruent spheres centered at the corresponding points as guidance, and then search for an increasing number of corresponding points. Since the image and 3D point cloud of the scene are both available from a Kinect device, we can readily obtain some 3D corresponding points according to the matched 2D SIFT-keypoints between images. Finally, the registration of point clouds can be achieved by applying least square to the searched corresponding points. The experimental results demonstrate the superior performance of the proposed method to the ICP algorithm.

Monocular 3D Tracking of Mean-Shift with Scale Adaptation Based on Projective Geometry

Mean-Shift a non-parameter algorithm, is widely used in real-time tracking. A crucial problem in implementing this algorithm is that there is presently no clear mechanism for updating the tracking window when a variable-size object is tracked. To solve this problem, we have proposed a 2D window updating algorithm based on projective geometry in the previous research.But on certain occasions, such as car tracking, a 3D window can describe the object more accurately than a 2D window. So we extend our algorithm into a 3D tracking, and the experiment shows that this algorithm performs well.

A Robust SIFT Method Based on Weighted Mean Phase Angles

The traditional interest points extraction, like SIFT, are based on the gradient operators. However, the gradient information is sensitive to image contrast, so it is important to select proper thresholds in the gradient-based method. To solve the problem, a robust SIFT method based on the weighted mean phase angles is proposed in this paper. The weighted mean phase angles are invariant to image contrast and illuminations, and could describe the feature types of points. Experiments show our method has a better performance than the traditional SIFT method.

Motion smoothness constraint for parameter optimization in real-time augmented reality

In professional video imaging, camera motion is always smooth, which is useful for the camera parameters optimization. Traditionally, optimization approaches for camera parameters focused on minimizing projection error. This paper presents a novel approach used in the real-time augmented reality software based on the Chebyshev Polynomial Fitting, using the smoothness of camera motion as an essential constraint. In order to characterize the performance of the approach, experiments with real-time video sequences are undertaken. The result analysis demonstrates the robustness of our optimization approach.

An adaptive early termination algorithm for motion estimation in multi-view video coding

In this paper, we propose an adaptive early termination algorithm for motion estimation in multi-view video coding (MVC). Currently, in order to find the optimal motion vector (MV) during MVC, all the reference frames need to be traversed, but the time consumption would be huge for a 7-modal block during the whole coding process. In the hope of reducing the coding time-consumption and meanwhile maintaining good video quality, we adopt adaptive termination thresholds based on each group of pictures (GOP) to early terminate the motion estimation (ME) process. Experimental results demonstrate that, compared with the Full Search (FS) and TZ-fast-search (TZFS) scheme in JMVC 7.2, the proposed algorithm reduces obviously the time-consumption in coding with negligible video-quality degradation.