PU-GCN: Point Cloud Upsampling using Graph Convolutional Networks

Qian, Guocheng; Abualshour, Abdulellah; Li, Guohao; Thabet, Ali; Ghanem, Bernard

doi:10.48550/arxiv.1912.03264

Cited by 7 publications

(14 citation statements)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To explore the efficacy of the proposed EVA upsampling unit, we conduct ablation study to compare with different upsampling units. Specifically, direct point feature duplication upsampling, and the best upsampling unit in PU-GCN [24], named NodeShuffle, are adopted for com-parison. The direct point feature duplication follows the network architecture of the Feature Expansion module in PU-Net [42], while keeps rest of network the same as PU-EVA.…”

Section: Ablation Studymentioning

confidence: 99%

PU-EVA: An Edge-Vector based Approximation Solution for Flexible-scale Point Cloud Upsampling

Luo

Tang

Zhou

et al. 2021

2021 IEEE/CVF International Conference on Computer Vision (ICCV)

View full text Add to dashboard Cite

High-quality point clouds have practical significance for point-based rendering, semantic understanding, and surface reconstruction. Upsampling sparse, noisy and nonuniform point clouds for a denser and more regular approximation of target objects is a desirable but challenging task. Most existing methods duplicate point features for upsampling, constraining the upsampling scales at a fixed rate. In this work, the flexible upsampling rates are achieved via edge vector based affine combinations, and a novel design of Edge Vector based Approximation for Flexible-scale Point clouds Upsampling (PU-EVA) is proposed. The edge vector based approximation encodes the neighboring connectivity via affine combinations based on edge vectors, and restricts the approximation error within the second-order term of Taylor's Expansion. The EVA upsampling decouples the upsampling scales with network architecture, achieving the flexible upsampling rates in one-time training. Qualitative and quantitative evaluations demonstrate that the proposed PU-EVA outperforms the state-of-the-arts in terms of proximity-to-surface, distribution uniformity, and geometric details preservation.

show abstract

Section: Ablation Studymentioning

confidence: 99%

PU-EVA: An Edge-Vector based Approximation Solution for Flexible-scale Point Cloud Upsampling

Luo

Tang

Zhou

et al. 2021

2021 IEEE/CVF International Conference on Computer Vision (ICCV)

View full text Add to dashboard Cite

show abstract

“…Specifically, unsupervised machine perception tasks extract informative features from point cloud samples without human supervision [9]. Some typical tasks include reconstruction [10], [11], completion [12], [13], and upsampling [14]. Recently, deep neural networks are common tools to achieve those tasks, and an objective point cloud distortion quantification is usually needed as the supervision to train deep neural networks.…”

Section: Machine Perception Tasksmentioning

confidence: 99%

“…Specifically, unsupervised tasks of machine perception generally aim to extract informative features from point clouds without using human supervision [9]. Some typical tasks include reconstruction [10], [11], completion [12], [13], and upsampling [14]. Recently, deep neural networks are emerging techniques to realize those tasks.…”

Section: Introductionmentioning

confidence: 99%

MPED: Quantifying Point Cloud Distortion based on Multiscale Potential Energy Discrepancy

Yang¹,

Zhang²,

Chen³

et al. 2021

Preprint

View full text Add to dashboard Cite

Distortion quantification of point clouds plays a stealth, yet vital role in a wide range of human and machine perception tasks. For human perception tasks, a distortion quantification can substitute subjective experiments to guide 3D visualization; while for machine perception tasks, a distortion quantification can work as a loss function to guide the training of deep neural networks for unsupervised learning tasks. To handle a variety of demands in many applications, a distortion quantification needs to be distortion discriminable, differentiable, and have a low computational complexity. Currently, however, there is a lack of a general distortion quantification that can satisfy all three conditions. To fill this gap, this work proposes multiscale potential energy discrepancy (MPED), a distortion quantification to measure point cloud geometry and color difference. The proposed MPED is able to capture both geometrical and color impairments by quantifying the total distortion between reference and distorted samples. By evaluating at various neighborhood sizes, the proposed MPED achieves global-local tradeoffs, capturing distortion in a multiscale fashion. We further theoretically show that classical Chamfer distance is a special case of our MPED. Extensive experimental studies validate MPED's superiority for both human and machine perception tasks. For human perception tasks, the proposed MPED works as subjective score predictor. In terms of Spearman rank-order correlation coefficient, MPED is 4% to 35% better than other state-of-the-art distortion quantifications on SJTU-PCQA database, and 27% to 190% on LSPCQA database. For machine perception tasks, the proposed MPED is plugged in as the loss function to enable the training of deep neural networks for three tasks, including point cloud reconstruction, shape completion and upsampling. The experimental results reveal that the proposed MPED produces better results than the point-wise Chamfer distance and Earth Mover's distance under the same network architecture. For instance, in point cloud reconstruction, MPED is over 80% and 70% better than Chamfer distance and Earth Mover's distance in terms of Jensen-Shannon divergence, respectively. We further study the robustness and convergence rate of MPED in ablation study, and the results show that: i) MPED is robust to the variations in color space, neighborhood scale and spatial field; ii) MPED can converge to stable results with less training time and epochs using the same network model. Our code is avaliable at https://github.com/Qi-Yangsjtu/MPED.

show abstract

“…To enhance the mesh resolution, we used a cascade of Multi-branch GCN [25] modules, where GCNConv layers [17] were used for feature upsampling. For the dome dataset we used 5 Multi-branch GCN modules at the first cascade level and then 8 modules in the second cascaded level and for the FreiHAND dataset we use 3 Multi-branch GCN modules.…”

Section: Mesh Enhancermentioning

confidence: 99%

“…For the dome dataset we used 5 Multi-branch GCN modules at the first cascade level and then 8 modules in the second cascaded level and for the FreiHAND dataset we use 3 Multi-branch GCN modules. The resultant node features which construct mesh at full resolution were then passed through a set of Convolution-BatchNorm-ReLU which plays a role analogues to the role of "Coordinator Reconstructor" of the initial work on point upsampling [25]. This contains Convolution-BatchNorm-ReLU layers with 64, 64, 64, 64 and 1 kernels with each kernel having 1 × 3 filters.…”

Section: Mesh Enhancermentioning

confidence: 99%

Im2Mesh GAN: Accurate 3D Hand Mesh Recovery from a Single RGB Image

Pemasiri¹,

Nguyen²,

Sridharan³

et al. 2021

Preprint

View full text Add to dashboard Cite

This work addresses hand mesh recovery from a single RGB image. In contrast to most of the existing approaches where the parametric hand models are employed as the prior, we show that the hand mesh can be learned directly from the input image. We propose a new type of GAN called Im2Mesh GAN to learn the mesh through end-to-end adversarial training. By interpreting the mesh as a graph, our model is able to capture the topological relationship among the mesh vertices. We also introduce a 3D surface descriptor into the GAN architecture to further capture the 3D features associated. We experiment two approaches where one can reap the benefits of coupled groundtruth data availability of images and the corresponding meshes, while the other combats the more challenging problem of mesh estimations without the corresponding groundtruth. Through extensive evaluations we demonstrate that the proposed method outperforms the state-of-the-art.

show abstract

PU-GCN: Point Cloud Upsampling using Graph Convolutional Networks

Cited by 7 publications

References 40 publications

PU-EVA: An Edge-Vector based Approximation Solution for Flexible-scale Point Cloud Upsampling

PU-EVA: An Edge-Vector based Approximation Solution for Flexible-scale Point Cloud Upsampling

MPED: Quantifying Point Cloud Distortion based on Multiscale Potential Energy Discrepancy

Im2Mesh GAN: Accurate 3D Hand Mesh Recovery from a Single RGB Image

Contact Info

Product

Resources

About