- Color/Depth image based object segmentation
We propose a robust edge refinement algorithm for object segmentation using a pair of color and depth images. We assume that the pixels belonging to a same object yield approximately similar depth values, and thus employ the gradient of depth image as a candidate for true object boundaries. Moreover, we also refine the initially obtained depth edges using the relatively detailed color edges by finding the best matching patterns between two edge images. Then, we use the refined depth edges to design the smoothness cost based on the Grabcut framework. The experiment results demonstrate that the proposed algorithm not only extracts the object boundaries accurately, but also reduces the false alarms of textured background.
- 3D geometry processing
In this work, we propose a geometry histogram modification algorithm to increase the visibility of 3D point models. We define a weak feature as a group of neighboring points yielding the small deviations of normal directions. Geometry histogram is defined as the distribution of the signed distance between a feature point and the locally approximated plane. We equalize and stretch the geometry histogram and move the corresponding feature points accordingly. We also use the OpenGL API for simple and fast rendering of 3D point models. Experimental results show that the proposed algorithm efficiently enhances the geometry contrast of 3D point models by improving the appearance of the weak features.