Graphical Models最新文献

3D models of trees are ubiquitous in video games, movies, and simulators. It is of paramount importance to generate high quality 3D models to enhance the visual content, and increase the diversity of the available models. In this work, we propose a methodology to create realistic 3D models of tree barks from a consumer-grade hand-held camera. Additionally, we present a pipeline that makes use of multi-view 3D Reconstruction and Generative Adversarial Networks (GANs) to generate the 3D models of the barks. We introduce a GAN referred to as the Depth-Reinforced-SPADE to generate the surfaces of the tree barks and the bark color concurrently. This GAN gives extensive control on what is being generated on the bark: moss, lichen, scars, etc. Finally, by testing our pipeline on different Northern-European trees whose barks exhibit radically different color patterns and surfaces, we show that our pipeline can be used to generate a broad panel of tree species’ bark.

Previous object-level Simultaneous Localization and Mapping (SLAM) approaches still fail to create high quality object-oriented 3D map in an efficient way. The main challenges come from how to represent the object shape effectively and how to apply such object representation to accurate online camera tracking efficiently. In this paper, we provide ObjectFusion as a novel object-level SLAM in static scenes which efficiently creates object-oriented 3D map with high-quality object reconstruction, by leveraging neural object priors. We propose a neural object representation with only a single encoder–decoder network to effectively express the object shape across various categories, which benefits high quality reconstruction of object instance. More importantly, we propose to convert such neural object representation as precise measurements to jointly optimize the object shape, object pose and camera pose for the final accurate 3D object reconstruction. With extensive evaluations on synthetic and real-world RGB-D datasets, we show that our ObjectFusion outperforms previous approaches, with better object reconstruction quality, using much less memory footprint, and in a more efficient way, especially at the object level.

The quasi-Bézier surface is a kind of commonly used surfaces in CAGD/CAD systems. In this paper, we present a novel approach to construct quasi-Bézier surfaces from the boundary information based on a general second order functional. This functional includes many common functionals as special cases, such as the Dirichlet functional, the biharmonic functional and the quasi-harmonic functional etc. The problem turns into solving simple linear equations about inner control points, and finally the internal control points of the resulting quasi-Bézier surface can be obtained as linear combinations of the given boundary control points. Some representative examples show the effectiveness of the presented method.

Point clouds are a widely used form of 3D data, which can be produced by depth sensors, such as RGB-D cameras. The classification of common elements of 3D point clouds remains an open research problem.

We propose a new deep network approach for the end-to-end training of log-Euclidean Fisher vectors (LE-FVs), applied to the classification of 3D point clouds. Our method uses a log-Euclidean (LE) metric in order to extend the concept of Fisher vectors (FVs) to LE-FV encoding. The LE-FV was computed on covariance matrices of local 3D point cloud descriptors, representing multiple features. Our architecture is composed of two blocks. The first one aims to map the covariance matrices representing the 3D point cloud descriptors to the Euclidean space. The second block allows for joint and simultaneous learning of LE-FV Gaussian Mixture Model (GMM) parameters, LE-FV dimensionality reduction, and multi-label classification.

Our LE-FV deep learning model is more accurate than the FV deep learning architecture. Additionally, the introduction of joint learning of 3D point cloud features in the log-Euclidean space, including LE-FV GMM parameters, LE-FV dimensionality reduction, and multi-label classification greatly improves the accuracy of classification. Our method has also been compared with the most popular methods in the literature for 3D point cloud classification, and it achieved good performance. The quantitative evidence will be shown through different experiments.

We introduce a novel method of isometric correspondences for 3D shapes, designed to address the problem of multiple solutions associated with deep functional maps when matching shapes with left-to-right reflectional intrinsic symmetries. Unlike the existing methods that only find the direct correspondences using single Siamese network, our proposed method is able to detect both the direct and symmetric correspondences among shapes simultaneously. Furthermore, our method detects the reflectional intrinsic symmetry of each shape. Key to our method is the using of two Siamese networks that learn consistent direct descriptors and their symmetric ones, combined with carefully designed regularized functional maps and supervised loss. This leads to the first deep functional map capable of both producing two high-quality correspondences of shapes and detecting the left-to-right reflectional intrinsic symmetry of each shape. Extensive experiments demonstrate that the proposed method obtains more accurate results than state-of-the-art methods for shape correspondences and reflectional intrinsic symmetries detection.

Mesh inpainting aims to fill the holes or missing regions from observed incomplete meshes and keep consistent with prior knowledge. Inspired by the success of low rank in describing similarity, we formulate the mesh inpainting problem as the low rank matrix recovery problem and present a patch-based mesh inpainting algorithm. Normal patch covariance is adapted to describe the similarity between surface patches. By analyzing the similarity of patches, the most similar patches are packed into a matrix with low rank structure. An iterative diffusion strategy is first designed to recover the patch vertex normals gradually. Then, the normals are refined by low rank approximation to keep the overall consistency and vertex positions are finally updated. We conduct several experiments in different 3D models to verify the proposed approach. Compared with existing algorithms, our experimental results demonstrate the superiority of our approach both visually and quantitatively in recovering the mesh with self-similarity patterns.

An accurate coarse-to-fine out-of-core outlier removal method is proposed for large-scale indoor point clouds by mining the geometric shape constraints. In coarse processing stage, a low-resolution point cloud (LPC) is obtained using random downsampling. LPC has the same density distribution as the raw point clouds (RPC), which is important information for outlier removal. The correspondences from the LPC to the RPC are also recorded. The outliers in the LPC are removed via a global threshold. The outliers in the RPC are roughly removed guided by the cleaned LPC. In refinement processing stage, the cleaned LPC is segmented into planar and non-planar segments; and the LPC segmentation is transferred to the RPC. Finally, the outliers in each RPC segment are removed elaborately via a local threshold by exploring the shape information. The experiments show that the proposed method improves the quality of outlier removal results.

We propose a novel method to untangle and optimize all-hex meshes. Central to this algorithm is an adaptive boundary optimization process that significantly improves practical robustness. Given an all-hex mesh with many inverted hexahedral elements, we first optimize a high-quality quad boundary mesh with a small approximation error to the input boundary. Since the boundary constraints limit the optimization space to search for the inversion-free meshes, we then relax the boundary constraints to generate an inversion-free all-hex mesh. We develop an adaptive boundary relaxation algorithm to implicitly restrict the shape difference between the relaxed and input boundaries, thereby facilitating the next step. Finally, an adaptive boundary difference minimization is developed to effectively and efficiently force the distance difference between the relaxed boundary and the optimized boundary of the first step to approach zero while avoiding inverted elements. We demonstrate the efficacy of our algorithm on a data set containing 1004 all-hex meshes. Compared to previous methods, our method achieves higher practical robustness.

The tooth axes, defined on 3D tooth model, play a key role in digital orthodontics, which is usually used as an important reference in automatic tooth arrangement and anomaly detection. In this paper, we propose an automatic deep learning network (TAD-Net) of tooth axis detection based on rotation transformation encoding. By utilizing quaternion transformation, we convert the geometric rotation transformation of the tooth axes into the feature encoding of the point cloud of 3D tooth models. Furthermore, the feature confidence-aware attention mechanism is adopted to generate dynamic weights for the features of each point to improve the network learning accuracy. Experimental results show that the proposed method has achieved higher detection accuracy on the constructed dental data set compared with the existing networks.

Recent advances in the design and fabrication of personalized figurines have made the creation of high-quality figurines possible for ordinary users with the facilities of 3D printing techniques. The hair plays an important role in gaining the realism of the figurines. Existing hair reconstruction methods suffer from the high demand for acquisition equipment, or the result is approximated very coarsely. Instead of creating hairs for figurines by scanning devices, we present a novel surface reconstruction method to generate a 3D printable hair model with geometric features from a strand-level hairstyle, thus converting the exiting digital hair database to a 3D printable database. Given a strand-level hair model, we filter the strands via bundle clustering, retain the main features, and reconstruct hair strands in two stages. First, our algorithm is the key to extracting the hair contour surface according to the structure of strands and calculating the normal for each vertex. Next, a close, manifold triangle mesh with geometric details and an embedded direction field is achieved with the Poisson surface reconstruction. We obtain closed-manifold hairstyles without user interactions, benefiting personalized figurine fabrication. We verify the feasibility of our method by exhibiting a wide range of examples.