Guimei Zhang

Disparity map optimization using sparse gradient measurement under intensity-edge constraints

Common local stereo methods often compute integer-valued disparities at support windows. The implicit assumption of a constant disparity value in support windows generally does not produce accurate results on slanted surfaces. In this paper, we propose a global optimization method for reconstructing disparity maps. Our optimization strategy is an extension of Xu’s (ACM Trans Graph 30(6):1–12, 2011) image smoothing algorithm. The strategy is to develop a sparse gradients counting model of the disparity map, coupled with the priors of intensity edges of the reference image. Based on this model, the disparity optimization problem is then formulated as a constrained optimization objective function, which is finally solved via the half-quadratic splitting algorithm. Experimental results demonstrated that the proposed approach improves the quality of the disparity maps regarding slanted surfaces and local discontinuity preservation.

Detection of saliency maximally stable color regions

In this study, we propose to detect regions of interest based on salient information in images. The maximally stable color region (MSCR) approach is extended by incorporating color salient information into the stable region detection design. Salient regions with a color similar to that of their vicinity are detected successively through agglomerative clustering. The algorithm introduces novel methods used in color saliency enhancement into the context of local feature detection. The color saliency enhancement approach is evaluated by detecting salient objects. Experimental results demonstrate that our algorithm yields high precision and recall rates, and focuses on the interesting color structure of the image. The proposed detector is also evaluated by using an image matching test. The experimental results show that this detector outperforms intensity- and color-based detectors in terms of match correspondence.

Color-based corner detection by color invariance

Gray-based features are widely used for computer vision applications, even when color data is a very important source. In this paper, we present a novel approach for detecting corners based on color information in images. To fully exploit the color data, a color invariant to saturation is first built based on dichromatic reflection model. The invariant is an object reflectance property independent of the viewpoint and illumination direction. Saturation information of the color is then synthesized with hue information to detect corners in color images. An extensive comparison of our detector with existing ones is carried out to validate a better performance in ground-truth verification and repeatability.

Recognizing a Planar Curve Based on NRLCTI and Area Threshold

In order to recognize that the curve has the same feature points but the curvature between the two adjacent feature points is different, a new method for matching the contour curve is presented in this chapter. First, the definition of NRLCTI (Normalized Run Length Code of Conner and Tangent and Inflexion Points) of a planar curve is given. In terms of NRLCTI, feature points both on object and models can be matched preliminarily. Subsequently, a new method is designed to match subcurves between the two adjacent feature points. We sample points on the subcurve based more on the precision requirement using the given minimal area threshold. A new recognition vector of sample points is defined, and a novel recognition vector matrix is constructed based on the recognition vector of sample points. Last, the dissimilarity measure of the corresponding subcurves is calculated by comparing the recognition vector matrix. The curve is recognized by matching all its subcurves. The method matches the object and model from simple to complex, so that many redundancies in calculations can be avoided. The experiment results show that the method is efficient and feasible.

Linear camera calibration and pose estimation from vanishing points

In order to broaden the scope of application and ensure the calibration precision, a new method of linear calibration by making fully use of vanishing point attributes is proposed. The method dose not need any rigorous restrictions, and solves the self-calibration problem with only five vanishing points in two digital images, which are arbitrarily taken by a handheld digital camera. Furthermore, another approach for camera’s pose estimation is also put forward without any strict controlled motions. The experimental results of both computer simulation and real images show that the calibration algorithm is effective, feasible, and robust. OCIS codes: 100.2000, 100.2960, 100.6890. doi: 10.3788/COL201210.S11007.

A novel global relaxation iteration algorithm for panoramic annular 3D terrain registration

It is necessary to reconstruct a large-scale landing-site mapping by recovering and registering the local scenes into a uniform annular scene for planetary exploration missions. This paper proposed a global relax iterative optimization method to registering the local scenes into a uniform annular scene. For this scheme, the transform matrix between any two adjacent 3D local scenes is fitted based on Carley transform. Subsequently, these local 3D scenes are registered into a uniform coordinate system using relax iterative optimization method. This optimization method has been tested on the image sequence of outdoor scenes. Experimental results show that the global registration means error decreases significantly from 1.33 meters to 0.002 meters in 47 images.

A wide baseline matching method based on scale invariant feature descriptor

Image matching is a fundamental aspect of many problems in computer vision. We describe a novel wide baseline matching method based on scale invariant feature descriptor. First, corners in image pairs are detected based on an improved Curvature Scale-Space (CSS) technique. These corners are relatively invariant to affine transformations, and are represented by using Scale Invariant Feature Transform (SIFT) descriptor to provide robust matching. The nearest neighbor distance is then applied to remove mismatched corners. Finally, the robust estimation algorithm, RANSAC, is adopt to estimate the fundamental matrix from the correspondence, and at the same time identify inlying matches. Experiments demonstrate the feasibility of this method.