DHGCN: Dynamic Hop Graph Convolution Network for Self-Supervised Point Cloud Learning

Introduction

• Existing point-cloud GCNs commonly build local graphs for grouped point sets and focus on local features, leaving contextual relationships between sets unmodeled.Point clouds’ irregular point distributions also make traditional image CNNs difficult to apply directly.Source: Introduction · pp. 1–2
• Graph hop distance provides a quantitative measure of adjacency between point parts, motivating its use to represent their contextual relationships.The paper’s figure constructs graph nodes from voxelized parts and depicts hop distance as shortest-path length.Source: Figure 1 · p. 1, Introduction · p. 2
• DHGCN voxelizes point clouds into graph nodes and reconstructs a part-level hop-distance matrix as a self-supervised task for learning contextual relationships.The dynamically updated distance matrix supervises learning and supplies geometric information to Hop Graph Attention, which weights edge features during aggregation.Source: Introduction · p. 2
• The paper contributes Hop Graph Attention and a hop-distance loss, and reports state-of-the-art performance across different downstream tasks.DHGCN is also designed as a plug-and-play module for point-based backbones.Source: Introduction · p. 2

Self-supervised Point Cloud Learning

• Point-cloud self-supervised methods use contrastive objectives across transformed views or reconstruct intra-cloud information from degraded inputs.Examples include matching point clouds to rendered images and completing occluded clouds.Source: Self-supervised Point Cloud Learning · p. 2

Graph-based Learning Methods

• Graph-based point-cloud methods commonly use kNN or radius neighborhoods and develop local feature extractors for the resulting point sets.The cited examples include DGCNN’s EdgeConv, deformable kernels, adaptive kernels, and deeper GCN architectures.Source: Graph-based Learning Methods · p. 2

Attention Mechanism

• Prior attention methods assign weights to graph neighbors or point features, whereas DHGCN’s HGA explicitly learns contextual relationships between point sets using geometric information.The paper motivates this approach through the adjacency structure among parts.Source: Attention Mechanism · pp. 2–3

Method

• DHGCN voxelizes the cloud into graph nodes, predicts their hop-distance matrix as a self-supervised signal, and uses that matrix in HGA to embed learned geometry into point features.The architecture extracts part features while predicting the distance matrix.Source: Method · p. 3
• The predicted hop distances represent the degree of adjacency between nodes and provide the contextual information that HGA embeds into point features.This connects the graph’s pairwise neighboring relationships to the point-feature representation.Source: Method · p. 3
• DHGCN extracts part features alongside distance prediction, using the predicted matrix as input to Hop Graph Attention.The passage identifies HGA as the stage that embeds learned geometric information into point features.Source: Method · p. 3

Volumetric Partition

• Volumetric partition maps each point to one of V = s^3 voxel parts, with each part storing the indices of its points.Each part’s axis-aligned bounding box is up-scaled by a factor of 1.2; an empty part receives a bounding box of volume 0.Source: Volumetric Partition · p. 3
• Two graph nodes connect when their bounding boxes intersect, and each node also has a self-loop.The graph distance is the shortest-path edge count; the maximum is truncated to s + 1 hops and self-distance is 0.Source: Volumetric Partition · p. 3
• The resulting distance matrix encodes pairwise adjacency degree through shortest-path hop counts.The maximum hop distance is truncated to s + 1.Source: Volumetric Partition · p. 3

Part Feature Extraction

• Learnable point-feature convolution maps point-cloud representations to N point-wise features H with C dimensions.Source: Part Feature Extraction · p. 3
• Permutation-invariant max pooling aggregates point-wise features into a global feature f_g.Source: Part Feature Extraction · p. 3
• DHGCN pools point-wise features within each part to form V part features F, with p_i containing that part’s point indices.Source: Part Feature Extraction · p. 3

Part Convolution

• DHGCN initializes a graph over parts and uses part-level graph convolution to extract local geometric representations from edge features.The edge features combine shape information with neighborhood feature differences; PartConv projects them to new edge features.Source: Part Convolution · p. 4

Hop Distance Reconstruction

• A self-supervised task predicts pairwise hop distances as δ + 1 classes, while each layer’s predicted distance matrix is supervised by the ground-truth matrix.The ground-truth graph uses voxel parts and their scaled axis-aligned bounding boxes to determine adjacency.Source: Hop Distance Reconstruction · p. 4, Figure 3 · p. 4

Hop Graph Attention

• Hop Graph Attention uses predicted hop distances to produce attention coefficients, giving edge features distinct contributions during aggregation.Its stated goal is to assign more weight to parts with fewer hops; normalized coefficients are compared across connected neighbors.Source: Hop Graph Attention · p. 4

Revising Point-wise Features

• DHGCN repeats each part-level feature for its points and adds it to the original point-wise features to form revised representations.Source: Revising Point-wise Features · p. 4
• The predicted hop distances align with parts’ geometric relationships, and revised features are more distinctive than DGCNN’s and more related to geometric distance and shape.Source: Revising Point-wise Features · p. 4, Figure 4 · p. 5

Dynamic Hop Distance

• Unlike local-graph methods, DHGCN uses a complete global graph whose hop distances are dynamically updated and supervised across layers.The Gaussian kernel assigns more attention to neighboring parts and less to distant parts; the self-supervised loss uses the HopGraphConv layers.Source: Part Convolution · p. 4, Dynamic Hop Distance · p. 5

Experimental Setting

• Classification uses split number 3 and maximum hop distance 4, while segmentation uses split number 5; HGA uses four attention heads and Gaussian-kernel σ is 1 by default.Experiments use SGD, batch size 32, and one TITAN RTX GPU.Source: Experimental Setting · p. 5

Pretraining

• DHGCN is pretrained on ShapeNet's 57,448 models across 55 categories and also on ModelNet40's 9,840 training models, then evaluated using a frozen encoder and linear classifier.ShapeNet pretraining uses 2,048 points; ModelNet40 pretraining uses 1,024 points.Source: Pretraining · p. 5

Downstream Tasks

• 93.2% (ShapeNet) and 93.3% (ModelNet40): DHGCN achieves state-of-the-art ModelNet40 classification, exceeding SSC by 0.8% and GraphTER by 1.3%, respectively.Table 1 compares unsupervised methods and identifies ShapeNet/ModelNet40 as the pretraining datasets.Source: Downstream Tasks · pp. 5–6, Table 1 · p. 6
• DHGCN achieves state-of-the-art results at all five ModelNet40 training-data ratios; with 1% data, it reaches 62.7% accuracy, exceeding MAE3D by 1% and FoldingNet by 6.3%.With 20% and 10% of training data, accuracy is 89.1% and 86.1%, respectively.Source: Downstream Tasks · p. 6, Table 2 · p. 6
• 85.9% on OBJ BG: DHGCN exceeds compared unsupervised methods by at least 4.2%; on PB T50 RS it reaches 81.9%, where the authors suspect perturbation noise disrupts part adjacency.On OBJ BG, it also surpasses the DGCNN backbone by 3.1%.Source: Downstream Tasks · p. 6, Table 3 · p. 6
• DHGCN achieves state-of-the-art performance over recent unsupervised methods on ShapeNet Part segmentation, evaluated with two mIoU types and PAConv as backbone.With limited training data, PAConv-backed DHGCN reaches 76.9% and 81.9% instance mIoU using 1% and 5% of training data, respectively.Source: Downstream Tasks · pp. 6–7, Table 4 · p. 7, Table 5 · p. 7

Ablation Studies

• 0.5%: HGA exceeds the supervised SA baseline; applying hop-distance loss in each layer yields 94.6% versus 93.3% hop prediction accuracy and 93.3% versus 93.1% classification accuracy.The table reports hop-distance prediction and point-cloud classification accuracy.Source: Ablation Studies · p. 7, Table 6 · p. 7
• 93.3% accuracy occurs at Gaussian-kernel σ² = 1.0; σ² = 0.2 reduces accuracy to 92.6%, while σ² = 2.0 and 5.0 yield 93.0% and 92.9%.The authors attribute the lower results at small σ² to a receptive field reduced to one-hop neighbors, and at larger σ² to less distance distinction.Source: Ablation Studies · p. 7, Table 7 · p. 7

• DHGCN outperforms state-of-the-art unsupervised methods on downstream classification and part segmentation tasks.Source: Conclusion · p. 7
• The self-supervised part-level hop-distance reconstruction task and hop-distance loss learn contextual relationships between point parts.Source: Conclusion · p. 7
• HGA embeds dynamically updated hop distances as attention weights to determine point parts’ importance during feature aggregation.Source: Conclusion · p. 7
• DHGCN is less effective on data with large perturbations, where noisy voxelization can produce misleading adjacent relationships.Source: Conclusion · p. 7