The authors propose a new non-rigid 3D shape retrieval method based on automatic anti-articulating. First, feature points are extracted on the 3D shape, and then are used in shape partition and skeleton extraction. Then, the mesh editing technique is employed to automatically regularize the articulation, in order to generate a rigid 3D shape without deformation, which is used as the isometry-invariant canonical form. Finally, shape descriptors are constructed on the canonical form, and Earth Mover’s distance is employed to measure the intrinsic similarities between non-rigid objects. The method introduces less geometric distortion and requires much less computational cost. Experimental results demonstrate a significant performance improvement in retrieving non-rigid 3D shapes.

The authors propose a new 3D shape representation, “local bag-of-words models”and use overlapped “local bag-of-words models”which improves the robustness of the algorithm by avoiding shape segmentation. The detailed steps are also proposed to apply the “Earth Mover’s distance”and “pyramid match kernels”algorithms to quantify partial shape similarity. It proves that the proposed distance measure mitigates the distance under-estimation problemof global bag-of-words model. Finally, the experiments firmly demonstratethe effectiveness of the proposed approach to 3Dpartial shape retrieval.

Focusing on scenes of outdoor buildings, the authors present graph-based stereo matching methods that incorporate depth cues to acquire more accurate disparity maps. Firstly, given a portion of scan data, planar disparity layers, which correspond to 3D planes, are precisely extracted to compose the label set. After that, the graph-based matching algorithmcan be formulated in the color segment domain instead of pixel domain. Furthermore,in order to avoid the inconvenient scanning for every scene, the scanner is only used to create a training set consisting of image-depthmap pairs. Then the depth cues, predicted through training a Markov Random Field to model the relationship between the depth and the image features, can still serve as extra constraint for the correspondence problem. The experimental results show the convincing performance of proposed methods in the reconstruction of different scenes.

A novel method is proposed to construct 3D shape descriptors by extracting rotation invariants of 3D models. In this method, a 3D model is firstly represented as a collection of spherical functions by using Hadamard transforms, and then the rotation invariants of the spherical functions are extracted by spherical harmonic decomposition. The method can avoid information, loss in the process of representing 3D models by the spherical functions. A shape similarity measure is defined on the extracted rotation invariants. A large amount of experiment results verified that this 3D shape descriptor performs better than some others.

Wavelete decomposition was performed to classify the source images into two types: the edge-riched image and the smooth image. Based on the Human Visual System theory, different combination techniques, the image stitching and the image blending methods, were selected to deal with different source images. A new algorithm was proposed to find an optimized stitching path in the overlapped area of Partial Texture Maps. View dependent weight was added into the ordinary blending weight, which made the generated texture map reveal some geometric properties of the 3D model. The experimental results verify the validity of the methods.

A new real-time omnidirectional vision system is proposed for tracking multiple targets in indoor environments. In this system, an omnidirectional camera is used to obtain the 360°view images of the global scene. First, it builds the background model by observing the scene without people in it. When the detecting process starts, omnidirectional images are changed into the cylindrical panoramic images, and an adaptive background subtraction method is utilized to segment the moving regions and locate the positions of human bodies. Thus, multiple human bodies can be tracked. Experimental results show that the proposed system performs well in indoor, complex environments.

The correspondence and branching problems are vexed problems when reconstructing 3-D surfaces from 2-D parallel contours. BPLI tries to reconstruct non-intersecting surfaces from arbitrary input set of contours and latch a lambent solution of the two problems; however, BPLI remains the shortages of producing degenerate portions and involving time-consuming operations. After an overall analysis of BPLI, this paper proposed a series of new algorithms to improve BPLI approach: first, we proposed a new contour matching algorithm so that the correct similar contour portions of each pair of adjacent slices can be found though the contours are not finely segmented. Secondly, we triangulate the unmatched portions with a new dynamic programming scheme, which has ability to handle degenerate portions in the mesh. These improvements result in a fast triangulating process.