Progressive Point Cloud Deconvolution Generation Network

1 Introduction

• Unordered, irregular point clouds make realistic point-cloud generation in an unsupervised way a challenging open problem.The paper notes that 2D generative models cannot be directly extended to point clouds.Source: §1 · p. 1
• Voxel-based 3D GANs face heavy computational and memory burdens, while point-cloud graph-convolution generators are costly to train.The cited prior work also includes a supervised, two-stage autoencoder-based model that is not end-to-end.Source: §1 · p. 2
• The proposed end-to-end network maps a latent vector through progressive deconvolution, using learning-based bilateral interpolation and MLPs to generate multi-resolution point clouds.Interpolation weights are learned from spatial and feature spaces, and the MLPs generate spatial coordinates.Source: §1 · p. 2
• A shape-preserving adversarial loss encourages point clouds generated from the same latent vector to retain consistent shapes across resolutions.The loss is introduced to train the progressive generation network.Source: §1 · p. 2

2.1 Deep Learning on 3D Data

• View-based and voxelized 3D methods apply standard CNNs, but voxel convolutions carry substantial memory and computational costs.The review also describes combining view-based and voxelization-based methods for shape classification.Source: §2.1 · p. 3
• Point-cloud methods include graph convolutions, PointNet, PointNet++ [28], dynamic graph CNN [39], and PointCNN, with PointNet++ designed to learn hierarchical features because PointNet captures local structures poorly.PointNet aggregates point features with MLPs and max-pooling.Source: §2.1 · p. 3

2.2 3D Point Cloud Generation

• VAE-based generation includes MRTNet and VAE/AAE approaches [46], while PointFlow models point clouds as a two-level hierarchical distribution.The passage notes that PointFlow is reported to converge slowly and fail on cases with many thin structures, such as chairs.Source: §2.2 · pp. 3–4
• Other point-cloud generators include the two-stage autoencoder/GAN model and graph-based localized or tree-structured generators.The review says the two-stage model is not end-to-end, while graph-convolution models can achieve impressive results.Source: §2.2 · p. 4

3.1 Progressive deconvolution generation network

• Learning-based bilateral interpolation uses feature-space k-NN neighborhoods and spatial and feature embeddings to exploit local geometric structure when enlarging feature maps.The embeddings use differences between points in the spatial and feature spaces to encode local structure.Source: §3.1 · pp. 5–6
• The deconvolution network uses MLPs and max pooling to form local features, then concatenates them with replicated global features to represent local geometry and overall shape.It divides each 2k-point neighborhood into the closest k points and the remaining k points before pooling local features.Source: §3.1 · pp. 6–7
• Stacked deconvolution networks progressively double the point count, while an MLP generates coordinates at each resolution and passes coordinates with features to the next block.The generator starts from a latent vector and builds high-dimensional feature maps.Source: §3.1 · pp. 4, 7
• PointNet discriminators distinguish generated from real point clouds, with different discriminator parameters for different resolutions.Source: §3.1 · p. 7

3.2 Shape-preserving adversarial loss

• The shape-preserving adversarial loss constrains adjacent-resolution outputs because unsupervised generators may produce different shapes for objects in the same class.It compares neighborhood means and covariances using Chamfer distances to encourage consistency between corresponding local structures.Source: §3.2 · pp. 8–9

4.2 Evaluation of point cloud generation

• Across four resolutions, generated shapes remain consistent, and PDGN produces more complete Airplane and Chair clouds than PointFlow, especially for thin structures.The visualizations cover 256, 512, 1024, and 2048 points.Source: §4.2 · pp. 10–12, Figure 4 · p. 10, Figure 5 · p. 12
• PDGN outperforms other methods on most reported criteria for the Airplane and Chair categories.Table 1 reports JSD, MMD, COV, and 1-NNA criteria; the caption states which metric directions indicate better performance.Source: Table 1 · p. 11, §4.2 · p. 11
• For 1,000 training iterations on 2,416 Airplane samples, PDGN takes about 1.9 days and 15G GPU memory, versus 4.5 days and 7.9G for PointFlow and 2.5 days and 9.2G for tree-GAN.The authors attribute PDGN's higher GPU memory use to its four discriminators.Source: §4.2 · p. 12
• On ModelNet10 and ModelNet40, PDGN outperforms state-of-the-art point-cloud generation methods in classification experiments.The evaluation extracts discriminator features and trains a linear SVM.Source: §4.2 · p. 12, Table 2 · p. 13

4.3 Ablation study and analysis

• Learning-based bilateral interpolation produces more realistic point clouds than bilinear interpolation or reshape, with more complete Table legs and the best Chair metrics.It learns weights from neighboring points' spatial coordinates and high-dimensional features.Source: Figure 6 · p. 13, §4.3 · p. 13, Table 3 · p. 14
• Shape-preserving loss maintains local geometric consistency across resolutions, whereas classical adversarial loss does not guarantee consistency and EMD or CD constraints preserve only global structure.The loss uses neighborhood means and covariances; Chair metrics are better than with the other tested losses.Source: §4.3 · p. 14, Table 3 · p. 14, Figure 7 · p. 15
• Bilateral interpolation performs better with k values around 20; very small neighborhoods weaken geometric features, while very large ones raise deconvolution cost.The authors test k from 2 to 36 in increments of 2.Source: §4.3 · p. 14, Figure 8 · p. 15
• As point-cloud resolution increases from 256 to 2048, reported quality improves and local structures become clearer.The progressive framework uses bilateral-interpolation-based deconvolution to build coarse-to-fine geometry.Source: §4.3 · p. 15

5 Conclusions

• The progressive deconvolution network generates multi-resolution point clouds from a latent vector, using bilateral interpolation to capture local structures during generation.The supplementary document outlines additional architecture, training, and visualization details.Source: §5 · p. 16
• A shape-preserving adversarial loss is formulated to keep the geometric structures of generated point clouds consistent across resolutions.Source: §5 · p. 16
• Experiments on ShapeNet and ModelNet verify the effectiveness of the proposed network.Source: §5 · p. 16

A Overview

• The supplementary document covers network architecture, training, additional visual and quantitative results, and comparisons with PointFlow.Source: A Overview · p. 18

B Network Architecture

• Four stacked deconvolution networks produce 256-, 512-, 1024-, and 2048-point outputs from a 128-dimensional latent vector.Each network has separate global-information and bilateral-interpolation branches; the networks are not shared across resolutions.Source: B Network Architecture · p. 18, Figure 10 · p. 19
• The four generator stages use MLPs to convert their feature maps into 3D coordinates, followed by Tanh activation.The figure labels G1 through G4 as MLPs for resolutions 256, 512, 1024, and 2048.Source: B Network Architecture · p. 18, Figure 10 · p. 19
• The method uses four PointNet-like discriminators, modified to accommodate different resolutions and not shared across resolutions; experiments found that too many convolution layers are harmful for low-resolution point clouds.Source: B Network Architecture · pp. 18–19

C Training of the Proposed GAN-based Model

• The model alternates one discriminator-optimization step with one generator-optimization step, using Adam at a learning rate of 10^-4 for both.The discriminator architecture uses separate networks for the four resolutions.Source: C Training of the Proposed GAN-based… · pp. 19–20, Figure 11 · p. 20
• The authors report stable training and quick convergence, suggesting progressive generation from low to high resolution may contribute to stability.They note that the low-resolution network is easier to train because it generates simpler shapes with fewer points, and its stability contributes to high-resolution network training.Source: C Training of the Proposed GAN-based… · p. 20

D.1 Generated multi-resolution point clouds

• Generated point clouds are reported as consistent across resolutions 256, 512, 1024, and 2048 across seven categories.The visualizations cover Airplane, Chair, Car, Table, Lamp, Pistol, and Guitar.Source: §D.1 · p. 20

D.2 Features in progressive deconvolution network.

• The feature-space analysis visualizes generated Airplane and Chair point clouds using outputs from four deconvolution networks and k-means clustering.Source: §D.2 · p. 20, Figure 12 · p. 21
• Chair parts become more clearly distinguished as the deconvolution network resolution increases.The feature-space visualizations use k-means clustering on Airplane and Chair outputs at resolutions 256, 512, 1024, and 2048.Source: §D.2 · p. 20, Figure 12 · p. 21

D.3 Three views of the generated point clouds

• Generated point clouds across seven categories are described as having realistic shapes compared with real point clouds.The categories are Airplane, Chair, Car, Table, Lamp, Pistol, and Guitar.Source: §D.3 · p. 21

• Across all 16 categories, PDGN performs better than r-GAN, tree-GAN, and PointFlow on the reported mean results.The authors report that they ran PointFlow’s official code to produce its mean results across all 16 categories.Source: §D.4 · p. 21
• On Car, PDGN’s metric results are comparable to PointFlow’s.Source: §D.4 · p. 21
• The authors suggest Car’s lack of many thin structures may explain its comparable results against PointFlow.Source: §D.4 · p. 21

• For thin-structured shapes such as chairs, PDGN performs well, while PointFlow is described as failing in such cases.The comparison uses PointFlow’s trained model provided on GitHub.Source: E Visualization Comparison with… · p. 21, Figure 22 · p. 27
• Figures 21 and 22 compare PDGN with PointFlow on Airplane and Chair point clouds at resolution 2048.Source: E Visualization Comparison with… · p. 21, Figure 21 · p. 26, Figure 22 · p. 27
• Figures 13–18 show visualization results for six categories—Airplane, Chair, Car, Table, Lamp, and Pistol—with each figure reporting resolutions 256, 512, 1024, and 2048, respectively.Source: Figure 13 · p. 22, Figure 14 · p. 22, Figure 15 · p. 23