Dual Octree Graph Networks for Learning Adaptive Volumetric Shape Representations

1 INTRODUCTION

• Regular-grid volumetric CNNs incur cubic memory and computation costs, while sparse-voxel methods ignore or zero empty-voxel features and make voxel-location prediction difficult.The latter difficulty is especially relevant for incomplete inputs.Source: §1 · p. 1
• Coordinate-based MLPs require many evaluations for high-resolution fields, while local-MLP detail fidelity remains limited by feature-volume resolution.The paper frames efficient volumetric representation and learning for shape generation and reconstruction as an open problem.Source: §1 · pp. 1–2
• An adaptive octree feature volume represents the full volumetric field, while an MLP maps its features and a queried 3D position to a field value.Coarse nodes represent low-frequency variation and fine nodes represent high-frequency detail around the surface; multilevel partition of unity combines neighboring-node contributions into a continuous field.Source: §1 · p. 2
• The method achieves state-of-the-art performance on four benchmarks and reconstructs surfaces about 390 times faster than MLP-based methods.The evaluation includes reconstruction from noisy point clouds, unsupervised surface reconstruction, and encoding 3D shape collections.Source: §1 · p. 2
• The encoder-decoder learns multiscale features from the input point cloud, and the new graph convolution improves convolution efficacy and computational and memory efficiency.The network contribution is described as learning shape details in high quality.Source: §1 · p. 2

2 RELATED WORK

• Regular-grid volumetric CNNs generalize 2D CNNs to 3D fields, but their computational and memory costs grow cubically with volumetric resolution.Sparse-voxel CNNs reduce these costs by operating on non-empty voxels.Source: §2 · p. 2
• Sparse-voxel approaches can omit empty-voxel features and often require ground-truth surfaces, limiting their use for unsupervised reconstruction from noisy or incomplete point clouds.The paper instead represents the full volumetric field with an adaptive octree and uses cross-level neighboring-node graph convolution.Source: §2 · p. 2
• Coordinate-based MLPs preserve field continuity and use memory efficiently, but high-resolution evaluation is costly and a single MLP struggles with rich detail or large scenes.Local-MLP methods reduce computation, while uniform feature grids can limit geometric detail and sparse-octree approaches may not guarantee continuity.Source: §2 · p. 3
• Existing graph CNNs for 3D commonly operate on point-cloud k-nearest-neighbor graphs, whereas this method operates on dual octree graphs and exploits their semi-regularity.Its operator is designed to improve graph-CNN efficiency and efficacy for this task.Source: §2 · p. 3
• The paper proposes a GPU scheme to construct multiple-level dual octree graphs, enabling its learning method to work across octree levels.The cited earlier dual-octree applications include isosurface extraction and surface reconstruction.Source: §2 · p. 3

3 DUAL OCTREE GRAPH NETWORKS

• The pipeline builds hierarchical dual octree graphs from an input point cloud, predicts an adaptive feature volume, and maps it to a continuous field with Neural MPU.The field can be an SDF or occupancy field, and the pipeline supports surface extraction.Source: §3 · pp. 3–4

3.1 Dual Octree Graph

• The dual octree graph connects face-adjacent leaf nodes, including nodes at different resolutions, while the octree allocates finer cells near the surface.The network requires graphs at multiple resolutions; it constructs them progressively by replacing subdivided vertices and their incident edges, then adding sibling edges.Source: §3.1 · p. 4

3.2 Dual Octree Graph CNNs

• The network operates on hierarchical dual octree graphs with graph convolution, downsampling, upsampling, and U-Net skip connections.In the decoder, predicted subdivision status can grow the octree when its subdivision is not known beforehand.Source: §3.2 · p. 4, §3.1 · p. 4

3.2.1 Dual octree graph convolution.

• The convolution uses a finite set of direction-specific weights, summing same-direction neighbors and augmenting features with node depth and relative position.The weight matrix has seven rows for the node and six directions; multiple neighbors in one direction share a weight. The implementation uses scatter and matrix multiplication, and the authors report it is more efficient and effective than prior KNN-based schemes.Source: §3.2.1 · p. 5

3.2.2 Other graph operations.

• Downsampling merges the finest sibling nodes into their parent, upsampling reverses that operation, and output-guided skips match encoder and decoder nodes by coordinates and depth.The skip connections support U-Nets when encoder and decoder octrees differ, as in shape completion.Source: §3.2.2 · pp. 5–6

3.3 Neural MPU

• Neural MPU blends node-local MLP predictions into a continuous, differentiable global field using locally supported weights that sum to one.Each local MLP takes a node feature and query-point coordinates; the method uses a single hidden layer and evaluates it after one network forward pass.Source: §3.3 · p. 6

3.4 Network and Loss functions

• The U-Net predicts octree subdivision and field values, and can be trained as a supervised or unsupervised network or as an autoencoder without skip connections.For surface reconstruction, the occupancy-field gradient loss constrains field values and gradients at input points and sampled volume points; the regression loss can also be applied at every octree layer.Source: §3.4 · pp. 6–7

4.1 Shape Reconstruction from Point Clouds

• The ShapeNet quantitative evaluation reports results on the testing set of 13 categories and separately on 5 unseen categories, with Chamfer distance multiplied by 100 for display.Source: Table 1 · p. 7
• On five unseen ShapeNet categories, our method again achieves the highest performance on all metrics, with results that change little from the test categories.The unseen categories are bathtub, bag, bed, bottle, and pillow.Source: Table 1 · p. 7, §4.1 · p. 8

4.2 Ablation Studies and Discussions

• FULLGRAPH improves over FINE→COARSE, which substantially outperforms SINGLESCALE and O-CNN; multiscale features and coarse-to-fine edges are important to performance.The authors conclude coarse features expand the receptive field and reduce high-frequency errors, while coarse-to-fine edges matter more than fine-to-coarse edges.Source: Table 2 · p. 8, §4.2 · p. 8
• Our graph convolution achieves the best performance and runs 5.1 times faster than KPConv and 5.5 times faster than EdgeConv.EdgeConv and KPConv were evaluated on the dual octree graph using the same training configurations.Source: Table 3 · p. 9, §4.2 · pp. 9–10
• Neural MPU produces continuous fields by interpolating features across resolutions, while the gradient loss is essential for reconstruction quality.Without the gradient term, metrics are CD: 0.269, NC: 0.949, IoU: 0.901, and F-Score: 0.990; performance drops slightly but remains above DeepMLS and ConvONet.Source: Table 1 · p. 7, §4.2 · p. 10
• The network has O(N^2) memory and computational cost and is reported as more efficient than volumetric CNNs, which run out of memory at resolution 64 in the ShapeNet setting.ConvONet uses 7.5G memory at resolution 32; DeepMLS and O-CNN use 9.4 GB and 8.7 GB, respectively, under the stated comparison settings.Source: §4.2 · p. 10, Table 4 · p. 10

4.3 Scene Reconstruction from Point Clouds

• On scene-level reconstruction, our method consistently outperforms the compared methods and is 2.7 times faster than ConvONet for the reported task.The method outputs adaptive grids up to resolution 128^3, whereas ConvONet uses dense voxels at resolution 64^3.Source: Table 5 · p. 11, §4.3 · p. 11

4.4 Unsupervised Surface Reconstruction

• On D-Faust, our unsupervised method outperforms IGR, with Chamfer distance one order smaller, and runs 394 times faster on average.It generalizes to Stanford scans and SceneNN scenes in a single forward pass, taking about 420 ms on average for the meshes in Fig. 11; 2,000 finetuning iterations further recover geometric detail.Source: Table 6 · p. 11, §4.4 · p. 12

4.5 Autoencoder

• The autoencoder is trained on 13 ShapeNet categories, using 35,019 training shapes and 8,762 test shapes, with octree depth 6.Chamfer distance and normal consistency use 4,096 sample points; all compared networks are trained on the 13 categories.Source: §4.5 · pp. 12–13
• Our method significantly outperforms IM-Net, OccNet, and AtlasNet and runs 5.9 times faster than IM-Net in the autoencoder experiment.The comparison uses Chamfer distance and normal consistency; running time measures a forward pass querying 100^3 uniformly sampled points per test shape.Source: §4.5 · pp. 12–13
• Compared with IM-Net, our reconstructions are more similar to ground truth, while IM-Net misses the chair base and airplane engine details.Figure 13 highlights artifacts and reduced detail in IM-Net results with red boxes.Source: §4.5 · p. 13, Figure 13 · p. 13

• Dual octree graph networks learn adaptive volumetric shape representations, using multiresolution graph convolution and blended features to generate continuous volumetric fields.The method partitions volume into adaptive voxels and uses encoder and decoder networks to extract adaptive features.Source: §5 · p. 13
• The current graph downsampling and upsampling merge or subdivide nodes at only one level at a time; simultaneous operations across leaf levels remain a possible alternative.The authors propose exploring whether this alternative could further maximize network performance.Source: §5 · p. 13
• Feature-volume adaptiveness depends on a prebuilt input octree that may not be optimal; combining octree optimization could increase compactness and efficiency.Source: §5 · p. 13
• On ModelNet40, the graph-convolution ResNet reaches 92.4% accuracy, below DGCNN (92.9%) and PCT (93.2%).Source: §5 · p. 13