Research map / Point clouds / Point cloud classification
DHGCN: Dynamic Hop Graph Convolution Network for Self-Supervised Point Cloud Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DHGCN, a plug-and-play graph module that self-supervises part-level hop-distance prediction over a complete graph of voxel parts and uses the predicted distances to weight edge features via Hop Graph Attention for self-supervised point cloud learning.
Problem
Existing GCN-based point cloud learners aggregate features within locally grouped point sets and ignore the contextual relationships between point sets; unsupervised point cloud representation learning needs better pretext tasks capturing this.
Approach
Voxelize the point cloud into parts, build a ground-truth graph whose edges connect intersecting scaled bounding boxes, and train the network to predict pairwise hop distances (a self-supervised task with a per-layer cross-entropy loss); the predicted distances then gate attention weights (via a Gaussian kernel) over edge features in a global complete part graph.
Key results
- Achieves 93.2% accuracy on ModelNet40 with ShapeNet pretraining and 93.3% with ModelNet40 pretraining under linear protocol, state of the art among unsupervised methods.
- Reaches 85.9% on ScanObjectNN OBJ_BG, exceeding unsupervised baselines by at least 4.2%.
- Obtains 82.9 class mIoU and 84.9 instance mIoU on ShapeNet Part, exceeding recent unsupervised methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DHGCN to the DGCNN backbone | DGCNN · ModelNet40 (ShapeNet pretraining) | accuracy: exceeding SSC by 0.8% | Helped |
| Add DHGCN to the AdaptConv backbone | AdaptConv · ModelNet40 (ModelNet40 pretraining) | accuracy: surpassing GraphTER by 1.3% | Helped |
| Use DHGCN rather than the strongest listed baseline | DGCNN · ModelNet40 with limited training data | accuracy: 1% data: 62.7% ... exceeding that of MAE3D by 1% | Helped |
| Use DHGCN rather than the strongest listed unsupervised baseline | DGCNN · ScanObjectNN OBJ_BG | accuracy: exceeding all compared SOTA unsupervised methods by at least 4.2% | Helped |
| Use DHGCN rather than the strongest listed unsupervised baseline | PAConv · ShapeNet Part | class mIoU | Helped |
| Use Hop Graph Attention instead of SA with λ=0 and no hop distance loss (ablation) | AdaptConv · ModelNet40 | accuracy: +0.5% | Helped |
| Calculate hop distance loss in every layer rather than only the last layer (ablation) | HGA with AdaptConv · ModelNet40 | distance accuracy: 93.3% versus 94.6% | Helped |
| Calculate hop distance loss in every layer rather than only the last layer (ablation) | HGA with AdaptConv · ModelNet40 | accuracy: 93.1% versus 93.3% | Helped |
| Set Gaussian kernel σ² to 0.2 rather than 1.0 (ablation) | HGA with AdaptConv · ModelNet40 | accuracy: 93.3% to 92.6% | Hurt |
| Set Gaussian kernel σ² to 0.5 rather than 1.0 (ablation) | HGA with AdaptConv · ModelNet40 | accuracy: 93.3% to 93.2% | Hurt |
| Set Gaussian kernel σ² to 2.0 rather than 1.0 (ablation) | HGA with AdaptConv · ModelNet40 | accuracy: 93.3% to 93.0% | Hurt |
| Set Gaussian kernel σ² to 5.0 rather than 1.0 (ablation) | HGA with AdaptConv · ModelNet40 | accuracy: 93.3% to 92.9% | Hurt |
Surprising findings
- Accuracy on the perturbed PB_T50_RS variant of ScanObjectNN (81.9%) is notably lower, which the authors attribute to voxel splitting errors under noise.
Limitations
- Less effective on data with large perturbations, since noise causes incorrect part splitting and misleading adjacency.
- Relies on voxel partition hyperparameters (split number s, max hop delta) chosen per task.
Experimental setup
Datasets: ShapeNet ModelNet40 ScanObjectNN ShapeNet Part
Benchmarks: ModelNet40 ScanObjectNN ShapeNet Part
Metrics: accuracy mIoU instance mIoU class mIoU
Baselines: CrossPoint SSC (RSCNN) GraphTER STRL VIPGAN MAE3D Point-MAE Point-BERT PAConv CTNet
Builds on: DGCNN PointNet++ AdaptConv GraphTER
Closest papers by meaning
- Improving Graph Representation for Point Cloud Segmentation via Attentive Filtering (CVPR 2023 · similarity 0.88)
- Learning Localized Generative Models for 3D Point Clouds via Graph Convolution (ICLR 2019 · similarity 0.88)
- Grid-GCN for Fast and Scalable Point Cloud Learning (CVPR 2020 · similarity 0.87)
- PU-GCN: Point Cloud Upsampling Using Graph Convolutional Networks (CVPR 2021 · similarity 0.87)
- Universal Graph Convolutional Networks (NeurIPS 2021 · similarity 0.87)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.