My main research interests are in image processing, 3D stereo matching, multi-resolution method, multi-array camera model.

Multi-array disparity enhancement

In this project, we use AUTODESK 3DS MAX software tool for 3D modeling and rendering. 3x3 array images and video are created for simulation. The initial disparity map is enhanced with aid of horizontal/vertical/narrow-baseline/wide-baseling disparity estimate. We propose a cascade Total Variation optimization to better refine object structure. The cascade TV can provide spatial, multi-array, temporal consistency. In addition, we propose a Cross-filling method to fuse different but complementary estimates. We obtain disparity enhancement by about 50% from the initial estimate.

3by3 model

Multi-resolution depth scheme for large panoramic view

Large stereo images seen on a virtual reality (VR) display are more favorable to customers because they can show realistic, high resolution imagery with a wide field of view. However, high resolution stereo images pose a challenging problem for many computer vision tasks. We present an effective multi-resolution depth processing for large stereo images. We propose an adaptively determined pixel-wise disparity search range, which is based on the combined eigenvalues of structure tensor matrix of image intensity and initial disparity. For the sub-pixel disparity, the multiple fitting algorithm is proposed to better represent the rounded surface while alleviating the pixel locking effect. To enforce the spatial and scaling consistency, we use the spatial-multiresolution TV method.

panoramic view

Local stereo matching algorithm

The resurgence of interest in 3D films and television has launched a new era in visual media consumption and research. Given a pair of stereoscopic views, disparity estimation is a crucial step for depth-based processing and communications. In the local stereo matching, the accuracy of the disparity map depends on the similarity measure and the support weight. We propose a novel three-moded census with a noise buffer to increase robustness to image noise in flat areas. We show that the combination of three similarity measures produces more reliable cost measure in a variety of image regions. To obtain more precise support weight, conditional and correlated support model are introduced. We consider object motion flow to take advantage of benefits of motion in video disparity estimation.

3-moded census