Graphical Models最新文献

Realistic 3D tree reconstruction is still a tedious and time-consuming task in the graphics community. In this paper, we propose a simple and efficient method for reconstructing 3D tree models with high fidelity from a single image. The key to single image-based tree reconstruction is to recover 3D shape information of trees via a deep neural network learned from a set of synthetic tree models. We adopted a conditional generative adversarial network (cGAN) to infer the 3D silhouette and skeleton of a tree respectively from edges extracted from the image and simple 2D strokes drawn by the user. Based on the predicted 3D silhouette and skeleton, a realistic tree model that inherits the tree shape in the input image can be generated using a procedural modeling technique. Experiments on varieties of tree examples demonstrate the efficiency and effectiveness of the proposed method in reconstructing realistic 3D tree models from a single image.

We present a new scale-adaptive ICP (Iterative Closest Point) method which aligns two objects that differ by rigid transformations (translations and rotations) and uniform scaling. The motivation is that input data may come in different scales (measurement units) which may not be known a priori, or when two range scans of the same object are obtained by different scanners. Classical ICP and its many variants do not handle this scale difference problem adequately. Our novel solution outperforms three different methods that estimate scale prior to alignment and a fourth method that, similar to ours, jointly optimizes for scale during the alignment.

Detecting ellipses from images is a fundamental task in many computer vision applications. However, due to the complexity of real-world scenarios, it is still a challenge to detect ellipses accurately and efficiently. In this paper, we propose a novel method to tackle this problem based on the fast computation of convex hull and directed graph, which achieves promising results on both accuracy and efficiency. We use Depth-First-Search to extract branch-free curves after adaptive edge detection. Line segments are used to represent the curvature characteristic of the curves, followed by splitting at sharp corners and inflection points to attain smooth arcs. Then the convex hull is constructed, together with the distance, length, and direction constraints, to find co-elliptic arc pairs. Arcs and their connectivity are encoded into a sparse directed graph, and then ellipses are generated via a fast access of the adjacency list. Finally, salient ellipses are selected subject to strict verification and weighted clustering. Extensive experiments are conducted on eight real-world datasets (six publicly available and two built by ourselves), as well as five synthetic datasets. Our method achieves the overall highest F-measure with competitive speed compared to representative state-of-the-art methods.

We introduce an A-weighting variance measurement, an objective estimation of the sound quality generated by geometric acoustic methods. Unlike the previous measurement, which applies to the impulse response, our measurement establishes the relationship between the impulse response and the auralized sound that the user hears. We also develop interactive methods to evaluate the measurement at run time and an adaptive algorithm that balances quality and performance based on the measurement. Experiments show that our method is more efficient in a wide variety of scene geometry, input sound, reverberation, and path tracing strategies.

The drone navigation requires the comprehensive understanding of both visual and geometric information in the 3D world. In this paper, we present a Visual-Geometric Fusion Network (VGF-Net), a deep network for the fusion analysis of visual/geometric data and the construction of 2.5D height maps for simultaneous drone navigation in novel environments. Given an initial rough height map and a sequence of RGB images, our VGF-Net extracts the visual information of the scene, along with a sparse set of 3D keypoints that capture the geometric relationship between objects in the scene. Driven by the data, VGF-Net adaptively fuses visual and geometric information, forming a unified Visual-Geometric Representation. This representation is fed to a new Directional Attention Model (DAM), which helps enhance the visual-geometric object relationship and propagates the informative data to dynamically refine the height map and the corresponding keypoints. An entire end-to-end information fusion and mapping system is formed, demonstrating remarkable robustness and high accuracy on the autonomous drone navigation across complex indoor and large-scale outdoor scenes.

Recent studies show increasing demands and interests in automatic layout generation, while there is still much room for improving the plausibility and robustness. In this paper, we present a data-driven layout generation framework without model formulation and loss term optimization. We achieve and organize priors directly based on samples from datasets instead of sampling probabilistic distributions. Therefore, our method enables expressing relations among three or more objects that are hard to be mathematically modeled. Subsequently, a non-learning geometric algorithm is proposed to arrange objects considering constraints such as positions of walls and windows. Experiments show that the proposed method outperforms the state-of-the-art and our generated layouts are competitive to those designed by professionals.¹

Unmanned airborne vehicles (UAVs) are useful in both military and civilian operations. In this paper, we consider a recreational scenario, i.e., multi-UAV formation transformation show. A visually smooth transformation needs to enforce the following three requirements at the same time: (1) visually pleasing contour morphing - for any intermediate frame, the agents form a meaningful shape and align with the contour, (2) uniform placement - for any intermediate frame, the agents are (isotropically) evenly spaced, and (3) smooth trajectories - the trajectory of each agent is as rigid/smooth as possible and completely collision free. First, we use the technique of 2-Wasserstein distance based interpolation to generate a sequence of intermediate shape contours. Second, we consider the spatio-temporal motion of all the agents altogether, and integrate the uniformity requirement and the spatial coherence into one objective function. Finally, the optimal formation transformation plan can be inferred by collaborative optimization.

Extensive experimental results show that our algorithm outperforms the existing algorithms in terms of visual smoothness of transformation, boundary alignment, uniformity of agents, and rigidity of trajectories. Furthermore, our algorithm is able to cope with some challenging scenarios including (1) source/target shapes with multiple connected components, (2) source/target shapes with different typology structures, and (3) existence of obstacles. Therefore, it has a great potential in the real multi-UAV light show. We created an animation to demonstrate how our algorithm works; See the demo at https://1drv.ms/v/s!AheMg5fKdtdugVL0aNFfEt_deTbT?e=le5poN .

Garment transfer from a source mannequin to a shape-varying individual is a vital technique in computer graphics. Existing garment transfer methods are either time consuming or lack designed details especially for clothing with complex styles. In this paper, we propose a data-driven approach to efficiently transfer garments between two distinctive bodies while preserving the source design. Given two sets of simulated garments on a source body and a target body, we utilize the deformation gradients as the representation. Since garments in our dataset are with various topologies, we embed cloth deformation to the body. For garment transfer, the deformation is decomposed into two aspects, typically style and shape. An encoder-decoder network is proposed to learn a shared space which is invariant to garment style but related to the deformation of human bodies. For a new garment in a different style worn by the source human, our method can efficiently transfer it to the target body with the shared shape deformation, meanwhile preserving the designed details. We qualitatively and quantitatively evaluate our method on a diverse set of 3D garments that showcase rich wrinkling patterns. Experiments show that the transferred garments can preserve the source design even if the target body is quite different from the source one.

Caricature is an artistic abstraction of the human face by distorting or exaggerating certain facial features, while still retains a likeness with the given face. Due to the large diversity of geometric and texture variations, automatic landmark detection and 3D face reconstruction for caricature is a challenging problem and has rarely been studied before. In this paper, we propose the first automatic method for this task by a novel 3D approach. To this end, we first build a dataset with various styles of 2D caricatures and their corresponding 3D shapes, and then build a parametric model on vertex based deformation space for 3D caricature face. Based on the constructed dataset and the nonlinear parametric model, we propose a neural network based method to regress the 3D face shape and orientation from the input 2D caricature image. Ablation studies and comparison with state-of-the-art methods demonstrate the effectiveness of our algorithm design. Extensive experimental results demonstrate that our method works well for various caricatures. Our constructed dataset, source code and trained model are available at https://github.com/Juyong/CaricatureFace.

Human motion transfer has many applications in human behavior analysis, training data augmentation, and personalization in mixed reality. We propose a Body-Parts-Aware Generative Adversarial Network (BPA-GAN) for image-based human motion transfer. Our key idea is to take advantage of the human body with segmented parts instead of using the human skeleton like most of existing methods to encode the human motion information. As a result, we improve the reconstruction quality, the training efficiency, and the temporal consistency via training multiple GANs in a local-to-global manner and adding regularization on the source motion. Extensive experiments show that our method outperforms the baseline and the state-of-the-art techniques in preserving the details of body parts.