About Me(CV)

Adaptive Normal Calculation(In Progress)

Normal calculation is not only important for model display but also vital for texture synthesis in 3D modeling. One way to calculate normal is based on gradient, classic marching cube normal estimation and 3D sobel operator are in this category. However, the distance function is relatively hard to calculate. Another way is based on plane fitting. For example, PlaneSVD, PlanePCA, and the like are in this category.

Actually, in our reconstruction algorithm, we already have indicator function. Our purpose is to estimate the normal accurately using known information.

Visual Hull Reconstruction Based on Adaptive Octree

In this Work, color images, silhouette images and calibration data are inputed. Tight spatial bounding box of model is first calculated, and then an adaptive octree is created based on flateness of original model. Finally, watertight triangle meshes are extracted from the adaptive octree.

Main contribution of this work: a new method of bounding box calculation; a node status(IN, ON and OUT which respectively represents that one node is totally inside of, intersect with, or totally out side of a model) checking algorithm; full adaptive octree based on flatness.

Details

Meanshift Segmentation Based On Color Constancy

Color is actually not an attribute that can be attached to the objects around us. It is basically a result of the processing done by the brain and the retina. The human visual system is able to determine the colors of objects irrespective of the illuminant. This ability is called Color Constancy.

In this work, we use retinex theory to correct the original color image and then segment the result image. The main purpose is to remove shadow and highlight from the original image and enhance the segmentation quality as a result.

Details

Stabilized K-means Clustering for Image Segmentation Based on GPU

In this work, we propose a new method of stable seeds selection in K-means clustering segmentation algorithm and the parallel implementation of this algorithm using GPU. Traditional K-means clustering method suffers from seeding problems and is slow when tackling a large amount of data. In our segmentation algorithm, we firstly get the histogram of one color component, and then choose the seeds by comparing the weight of each local peak, which is defined as the total number of pixels between the two adjacent valleys. We also port the stabilized K-means clustering segmentation algorithm to GPU for further acceleration and achieve satisfactory experimental results.

Details