Computer Vision and Computer Graphics. Theory and Applications

This paper addresses the problem of reconstructing implicit function from point clouds with noise and outliers acquired with 3D scanners. We introduce a filtering operator based on mean shift scheme, which shift each point to local maximum of kernel density function, resulting in suppression of noise with different amplitudes and removal of outliers. The “clean” data points are then divided into subdomains using an adaptive octree subdivision method, and a local radial basis function is constructed at each octree leaf cell. Finally, we blend these local shape functions together with partition of unity to approximate the entire global domain. Numerical experiments demonstrate robust and high quality performance of the proposed method in processing a great variety of 3D reconstruction from point clouds containing noise and outliers.

We consider the problem of extracting morphology of a terrain represented as a Triangulated Irregular Network (TIN). We propose a new algorithm and compare it with representative algorithms of the main approaches existing in the literature to this problem. The new algorithm has the advantage of being simple, using only comparisons (and no floating-point computations), and of being suitable for an extension to higher dimensions. Our experiments consider both real data and artificial test data. We evaluate the difference in the results produced on the same terrain data, as well as the impact of resolution level on such a difference, by considering representations of the same terrain at different resolutions.

We present a method for creating geometric models of dendritic forms. Dendritic shapes are commonplace in the natural world; some examples of objects exhibiting dendritic shape include lichens, coral, trees, lightning, rivers, crystals, and venation patterns. Our method first generates a regular lattice with randomly weighted edges, then finds least-cost paths through the lattice. Multiple paths from a single starting location (or generator) are connected into a single dendritic shape. Alternatively, path costs can be used to segment volumes into irregular shapes. The pathfinding process is inexpensive, and allows user control through specification of endpoint placement, distribution of generators, and arrangement of nodes in the graph.

This paper presents an integrated approach for interactive direct volume deformation and simultaneous visualization. The fundamental requirement is that interactive performance without pre-processing must be achieved for large volume data, where at any time up to one million elements participate in a deformation that is applied interactively by picking and dragging in the 3D view. Current physically-based approaches are still one or two orders of magnitude away from this goal. In contrast, our approach extends the non-physical ChainMail algorithm and combines it with on-the-fly resampling and GPU ray-casting. Special transfer functions assign material properties depending on volume density. The affected subvolume is deformed and resampled onto a rectilinear grid on the CPU, and updates the volume on the GPU where it is rendered using ray-casting. While the deformation is already being displayed, its quality is simultaneously refined via an iterative relaxation procedure executed in a parallel thread.

This paper presents a method for physical simulation of deformable closed surfaces over a network, which is suitable for realistic interactions between users and objects in a Collaborative Virtual Environment (CVE). To demonstrate a deformable object in a CVE, we employ a real-time physical simulation of a uniform-tension-membrane, based on linear finite-element-discretization of the surface. The proposed method introduces an architecture that distributes the computational load of physical simulation between each participant. Our approach requires a uniform-mesh representation of the simulated structure; therefore we designed and implemented a re-meshing algorithm that converts irregularly triangulated genus zero surfaces into a uniform triangular mesh with regular connectivity. The strength of our approach comes from the subdivision methodology that enables to use multi-resolution surfaces for graphical representation, physical simulation, and network transmission, without compromising simulation accuracy and visual quality.

Meshless deformation based on shape matching is a new technique for simulating deformable objects without requiring mesh connectivity information. The approach focuses on speed, ease of use and stability at the expense of physical accuracy. In this paper we introduce improvements to the technique that increase physical realism and make it more suitable for use in interactive real-time environments such as games and virtual surgery applications. We also present intuitive real-time interaction techniques for picking, pushing and cutting objects simulated using meshless deformation based on shape matching. For deformable collision detection and response, we present a new method for surface meshes based on previous volumetric methods.

Image stitching is an image processing method, where multiple photographs covering different parts of the same scene, are combined to form a single wide-angle image. Stitching is a very challenging task, and during the past decades many algorithms have been developed for it. Unfortunately, there has been no objective way to measure the quality of stitching results. To mend this shortcoming, we propose a novel method for testing stitching algorithms. The testing process starts from an arbitrary reference image that is used to create synthetic input data for the stitching algorithm that is to be tested. To make the testing realistic, various camera-related distortions along with perspective warps are applied to the input images. From this input data, the stitching algorithm creates a wide-angle image that is then compared to the reference image, from which the process started.

In this paper, we address the problem of robustly recovering several instances of a curve model from a single noisy data set with outliers. Using M-estimators revisited in a Lagrangian formalism, we derive an algorithm that we call Simultaneous Multiple Robust Fitting (SMRF), which extends the classical Iterative Reweighted Least Squares algorithm (IRLS). Compared to the IRLS, it features an extra probability ratio, which is classical in clustering algorithms, in the expression of the weights. Potential numerical issues are tackled by banning zero probabilities in the computation of the weights and by introducing a Gaussian prior on curves coefficients. Applications to camera calibration and lane-markings tracking show the effectiveness of the SMRF algorithm, which outperforms classical Gaussian mixture model algorithms in the presence of outliers.

As part of our work on hand gesture interpretation, we present our results on hand shape recognition. Our method is based on attribute extraction and multiple partial classifications. The novelty lies in the fashion the fusion of all the partial classification results are performed. This fusion is (1) more efficient in terms of information theory and leads to more accurate results, (2) general enough to allow heterogeneous sources of information to be taken into account: Each classifier output is transformed to a belief function, and all the corresponding functions are fused together with other external evidential sources of information.

We study the human action recognition problem based on motion features directly extracted from video. In order to implement a fast human action recognition system, we select simple features that can be obtained from non-intensive computation. We propose to use the motion history image (MHI) as our fundamental representation of the motion. This is then further processed to give a histogram of the MHI and the Haar wavelet transform of the MHI. The combination of these two features is computed cheaply and has a lower dimension than the original MHI. The combined feature vector is tested in a Support Vector Machine (SVM) based human action recognition system and a significant performance improvement has been achieved. The system is efficient to be used in real-time human action classification systems.

An integral component of face processing research is estimation of head orientation from face images. Head pose estimation bears importance in several applications in biometrics, human-computer interfaces, driver monitoring systems, video conferencing and social interaction enhancement programs. A recent trend in head pose estimation research has been the use of manifold learning techniques to capture the underlying geometry of the images. Face images with varying pose angles can be considered to be lying on a smooth low-dimensional manifold in high-dimensional image feature space. However, with real-world images, manifold learning techniques often fail because of their reliance on a geometric structure, which is often distorted due to noise, illumination changes and other variations. Also, when there are face images of multiple individuals with varying pose angles, manifold learning techniques often do not give accurate results. In this work, we introduce the formulation of a novel framework for supervised manifold learning called Biased Manifold Embedding to obtain improved performance in person-independent head pose estimation. While this framework goes beyond pose estimation, and can be applied to all regression applications, this work is focused on formulating the framework and validating its performace using the Isomap technique for head pose estimation. The work was carried out on face images from the FacePix database, which contains 181 face images each of 30 individuals with pose angle variations at a granularity of 1 ^∘. A Generalized Regression Neural Network (GRNN) was used to learn the non-linear mapping, and linear multi-variate regression was adopted on the low-dimensional space to obtain the pose angle. Results showed that the approach holds promise, with estimation errors substantially lower than similar efforts in the past using manifold learning techniques for head pose estimation.

We present a complete real-time model-based tracking system for piecewise-planar objects which combines template-based and feature-based approaches. Our contributions are an extension to the ESM algorithm used for template-based tracking and the formulation of a feature-based tracking approach, which is specifically tailored for use in a real-time setting. In order to cope with highly dynamic scenarios, such as illumination changes, partial occlusions and fast object movement, the system adaptively switches between template-based tracking, feature-based tracking and a global initialization phase. Our tracking system achieves real-time performance by applying a coarse-to-fine optimization approach and includes means to detect a loss of track.

In this paper we present a new approach that uses local structural information to find correspondences between image and model contour information. For a monocular pose estimation scenario, the pose is computed by our purposed new variant of the ICP (iterative closest point) algorithm which combines Euclidean distance with structure. A local representation of 3D free-form contours is used to get the structural information in 3D space and in the image plane. Furthermore, the local structure of free-form contours is combined with local orientation and phase obtained from the monogenic signal. With this combination, we achieve a more robust correspondence search. Our approach was tested on synthetical and real data to compare the convergence and performance of our approach against the classical ICP approach.

This paper presents a video-based camera tracker that combines mark- er-based and feature point-based cues in a particle filter framework. The framework relies on their complementary performance. Marker-based trackers can robustly recover camera position and orientation when a reference (marker) is available, but fail once the reference becomes unavailable. On the other hand, feature point tracking can still provide estimates given a limited number of feature points. However, these tend to drift and usually fail to recover when the reference reappears. Therefore, we propose a combination where the estimate of the filter is updated from the individual measurements of each cue. More precisely, the marker-based cue is selected when the marker is available whereas the feature point-based cue is selected otherwise. Feature points are dynamically found in scene and used for further tracking. Evaluations on real cases show that the fusion of these two approaches outperforms the individual tracking results. A critical aspect of the feature point-based cue is to robustly recognise the feature points depite rotations of the camera. A novelty of the proposed framework is the use of a rotation-discriminative method to match feature points.