Graphical Models最新文献

This paper aims at extending the method of Zhang et al. (2023) to produce not only portrait bas-reliefs from single photographs, but also high-depth reliefs with reasonable depth ordering. We cast this task as a problem of style-aware photo-to-depth translation, where the input is a photograph conditioned by a style vector and the output is a portrait relief with desired depth style. To construct ground-truth data for network training, we first propose an optimization-based method to synthesize high-depth reliefs from 3D portraits. Then, we train a normal-to-depth network to learn the mapping from normal maps to relief depths. After that, we use the trained network to generate high-depth relief samples using the provided normal maps from Zhang et al. (2023). As each normal map has pixel-wise photograph, we are able to establish correspondences between photographs and high-depth reliefs. By taking the bas-reliefs of Zhang et al. (2023), the new high-depth reliefs and their mixtures as target ground-truths, we finally train a encoder-to-decoder network to achieve style-aware relief modeling. Specially, the network is based on a U-shaped architecture, consisting of Swin Transformer blocks to process hierarchical deep features. Extensive experiments have demonstrated the effectiveness of the proposed method. Comparisons with previous works have verified its flexibility and state-of-the-art performance.

Image enhancement is an indispensable pre-processing step for several image processing applications. Mainly, histogram equalization is one of the widespread techniques used by various researchers to improve the image quality by expanding the pixel values to fill the entire dynamic grayscale. It results in the visual artifact, structural information loss near edges due to the information loss (due to many-to-one mapping), and alteration in average luminance to a higher value. This paper proposes an enhancement algorithm based on structural information for homogeneous background images. The intensities are divided into two segments using the median value to preserve the average luminance. Unlike traditional techniques, this algorithm incorporates the spatial locations in the equalization process instead of the number of intensity values occurrences. The occurrences of each intensity concerning their spatial locations are combined using Rènyi entropy to enumerate a discrete function. An adaptive clipping limit is applied to the discrete function to control the enhancement rate. Then histogram equalization is performed on each segment separately, and the equalized segments are integrated to produce an enhanced image. The algorithm’s effectiveness is validated by evaluating the proposed method on CEED, CSIQ, LOL, and TID2013 databases. Experimental results reveal that the proposed method improves the contrast while preserving structural information, detail information, and average luminance. They are quantified by the high value of contrast improvement index, structural similarity index, and discrete entropy, and low value of average mean brightness error values of the proposed method when compared with the methods available in the literature, including deep learning architectures.

This paper presents performance improvements for Ellipsoid Clipmaps, an out-of-core planet-sized geodetically accurate terrain rendering algorithm. The performance improvements were achieved by eliminating unnecessarily dense levels, more accurate block culling in the geographic coordinate system, and more efficient rendering methods. The elimination of unnecessarily dense levels is the result of analyzing and determining the optimal relative height of the viewer with respect to the most detailed level, resulting in the most consistent size of triangles across all visible levels. The proposed method for estimating the visibility of blocks based on view orientation allows rapid block-level view frustum culling performed in data space before visualization and spatial transformation of blocks. The use of a modern geometry pipeline through task and mesh shaders forced the handling of extremely fine granularity of blocks, but also shifted a significant part of the block culling process from CPU to the GPU. The approach described achieves high throughput and enables geodetically accurate rendering of the terrain based on the WGS 84 reference ellipsoid at very high resolution and in real time, with tens of millions of triangles with an average area of about 0.5 pix${}^2$ on a 1080p screen on mid-range graphics cards.

Point cloud upsampling has been extensively studied, however, the existing approaches suffer from the losing of structural information due to neglect of spatial dependencies between points. In this work, we propose PU-GAT, a novel 3D point cloud upsampling method that leverages graph attention networks to learn structural information over the baselines. Specifically, we first design a local–global feature extraction unit by combining spatial information and position encoding to mine the local spatial inter-dependencies across point features. Then, we construct an up-down-up feature expansion unit, which uses graph attention and GCN to enhance the ability of capturing local structure information. Extensive experiments on synthetic and real data have shown that our method achieves superior performance against previous methods quantitatively and qualitatively.

This paper presents a fast and progressive polygonal approximation algorithm for online strokes. A stroke is defined as a sequence of points between a pen-down and a pen-up. The proposed method generates polygonal approximations progressively as the user inputs the stroke. The proposed algorithm is suitable for real time shape modeling and retrieval. The number of operations used in the proposed algorithm is bounded by O(n), where n is the number of points in a stroke. Detailed experimental results show that the proposed method is not only fast, but also accurate enough compared to other reported algorithms.

We introduce a neural network, MixNet, for learning implicit representations of 3D subtle models with large smooth areas and exact shape details in the form of interpolation of two different implicit functions. Our network takes a point cloud as input and uses conventional MLP networks and SIREN networks to predict different implicit fields. We use a learnable interpolation function to combine the implicit values of these two networks and achieve the respective advantages of them. The network is self-supervised with only reconstruction loss, leading to faithful 3D reconstructions with smooth planes, correct details, and plausible spatial partition without any ground-truth segmentation. We evaluate our method on ABC, the largest and most diverse CAD dataset, and some typical shapes to test in terms of geometric correctness and surface smoothness to demonstrate superiority over current alternatives suitable for shape reconstruction.

In traditional deep functional maps for non-rigid shape correspondence, estimating a functional map including high-frequency information requires enough linearly independent features via the least square method, which is prone to be violated in practice, especially at an early stage of training, or costly post-processing, e.g. ZoomOut. In this paper, we propose a novel method called RFMNet (Robust Deep Functional Map Networks), which jointly considers training stability and more geometric shape features than previous works. We directly first produce a pointwise map by resorting to optimal transport and then convert it to an initial functional map. Such a mechanism mitigates the requirements for the descriptor and avoids the training instabilities resulting from the least square solver. Benefitting from the novel strategy, we successfully integrate a state-of-the-art geometric regularization for further optimizing the functional map, which substantially filters the initial functional map. We show our novel computing functional map module brings more stable training even under encoding the functional map with high-frequency information and faster convergence speed. Considering the pointwise and functional maps, an unsupervised loss is presented for penalizing the correspondence distortion of Delta functions between shapes. To catch discretization-resistant and orientation-aware shape features with our network, we utilize DiffusionNet as a feature extractor. Experimental results demonstrate our apparent superiority in correspondence quality and generalization across various shape discretizations and different datasets compared to the state-of-the-art learning methods.

To deal with the exhausting annotations, self-supervised representation learning from unlabeled point clouds has drawn much attention, especially centered on augmentation-based contrastive methods. However, specific augmentations hardly produce sufficient transferability to high-level tasks on different datasets. Besides, augmentations on point clouds may also change underlying semantics. To address the issues, we propose a simple but efficient augmentation fusion contrastive learning framework to combine data augmentations in Euclidean space and feature augmentations in feature space. In particular, we propose a data augmentation method based on sampling and graph generation. Meanwhile, we design a data augmentation network to enable a correspondence of representations by maximizing consistency between augmented graph pairs. We further design a feature augmentation network that encourages the model to learn representations invariant to the perturbations using an encoder perturbation. We comprehensively conduct extensive object classification experiments and object part segmentation experiments to validate the transferability of the proposed framework. Experimental results demonstrate that the proposed framework is effective to learn the point cloud representation in a self-supervised manner, and yields state-of-the-art results in the community. The source code is publicly available at: https://github.com/VCG-NJUST/AFSRL.

This paper presents a novel approach for text/speech-driven animation of a photo-realistic head model based on blend-shape geometry, dynamic textures, and neural rendering. Training a VAE for geometry and texture yields a parametric model for accurate capturing and realistic synthesis of facial expressions from a latent feature vector. Our animation method is based on a conditional CNN that transforms text or speech into a sequence of animation parameters. In contrast to previous approaches, our animation model learns disentangling/synthesizing different acting-styles in an unsupervised manner, requiring only phonetic labels that describe the content of training sequences. For realistic real-time rendering, we train a U-Net that refines rasterization-based renderings by computing improved pixel colors and a foreground matte. We compare our framework qualitatively/quantitatively against recent methods for head modeling as well as facial animation and evaluate the perceived rendering/animation quality in a user-study, which indicates large improvements compared to state-of-the-art approaches.

In this paper, we present an inverse rendering method for the simple reconstruction of shape and appearance of real-world objects from only roughly calibrated RGB images captured under collocated point light illumination. To this end, we gradually reconstruct the lower-frequency geometry information using automatically generated occupancy mask images based on a visual hull initialization of the mesh, to infer the object topology, and a smoothness-preconditioned optimization. By combining this geometry estimation with learning-based SVBRDF parameter inference as well as intrinsic and extrinsic camera parameter refinement in a joint and unified formulation, our novel method is able to reconstruct shape and an isotropic SVBRDF from fewer input images than previous methods. Unlike in other works, we also estimate normal maps as part of the SVBRDF to capture and represent higher-frequency geometric details in a compact way. Furthermore, by regularizing the appearance estimation with a GAN-based SVBRDF generator, we are able to meaningfully limit the solution space. In summary, this leads to a robust automatic reconstruction algorithm for shape and appearance. We evaluated our algorithm on synthetic as well as on real-world data and demonstrate that our method is able to reconstruct complex objects with high-fidelity reflection properties in a robust way, also in the presence of imperfect camera parameter data.