DHGCN: Dynamic Hop Graph Convolution Network for Point Cloud Learning (AAAI 2024)

Jincen Jiang, Lizhi Zhao, Xuequan Lu, Wei Hu, Imran Razzak, and Meili Wang.

[paper] [poster]

This is the source code for the implementation of our DHGCN for point cloud learning.

Introduction

Abstract

Recent works attempt to extend Graph Convolution Networks (GCNs) to point clouds for classification and segmentation tasks. These works tend to sample and group points to create smaller point sets locally and mainly focus on extracting local features through GCNs, while ignoring the relationship between point sets. In this paper, we propose the Dynamic Hop Graph Convolution Network (DHGCN) for explicitly learning the contextual relationships between the voxelized point parts, which are treated as graph nodes. Motivated by the intuition that the contextual information between point parts lies in the pairwise adjacent relationship, which can be depicted by the hop distance of the graph quantitatively, we devise a novel self-supervised part-level hop distance reconstruction task and design a novel loss function accordingly to facilitate training. In addition, we propose the Hop Graph Attention (HGA), which takes the learned hop distance as input for producing attention weights to allow edge features to contribute distinctively in aggregation. Eventually, the proposed DHGCN is a plug-and-play module that is compatible with point-based backbone networks. Comprehensive experiments on different backbones and tasks demonstrate that our self-supervised method achieves state of-the-art performance.

Highlight

Ours: Constructing the ground truth graph for our self-supervised hop distance reconstruction task. (a): Voxelizing the point cloud into parts, taking each part as a graph node. (b): The topology of the ground truth graph. Two nodes are adjacent if their scaled bounding boxes are intersected. (c): The shortest path between a node (enlarged red point) and other nodes. The number on each node denotes the hop distance which motivates our self-supervised task.

Others: Sampling and grouping based strategy. (d): Sampling center points and grouping local point sets. (e): Constructing a local graph for each point set. We explore the contextual relationships between parts, while previous strategies focus on extracting local features of point sets.

Network

Dynamic Update

Requirements

• CUDA = 11.3
• Python = 3.10
• PyTorch = 1.12.1
• Numpy = 1.23.1
• Package: h5py, sklearn, plyfile, json, cv2, yaml, tqdm

Different backbone networks for different tasks may have different requirements.

Contents

We provide source codes for different point cloud learning tasks:

• 3D Object Classification on ModelNet40 and ScanObjectNN:

  • Multi-task learning with different backbones (DGCNN, AdaptConv and PRA-Net) on ModelNet40.
  • Multi-task learning with DGCNN as the backbone on ScanObjectNN.
  • Unsupervised evaluation with linear classifier on ModelNet40.

• Shape Part Segmentation on ShapeNet Part:

  • Multi-task learning with different backbones (DGCNN and PAConv).
  • Unsupervised evaluation with PAConv as the backbone.

• Scene Segmentation on S3DIS.

Please refer to the directory of each task for the usage details of our code.

Results

3D Object Classification

• Multi-task Learning on ModelNet40

Methods	Accuracy
PRA-Net	93.2
DHGCN (PRA-Net)	93.4 (+0.2)
DGCNN	92.9
DHGCN (DGCNN)	93.5 (+0.6)
AdaptConv	93.5
DHGCN (AdaptConv)	93.6 (+0.2)

• Multi-task Learning on ScanObjectNN

Methods	OBJ_ONLY	OBJ_BG	PB_T50_RS
DGCNN	86.2	82.8	78.2
DHGCN (DGCNN)	88.3 (+2.1)	88.3 (+5.5)	82.9 (+4.8)

• Unsupervised Representation Learning on ModelNet40

Methods	Pretrained dataset	Accuracy
DHGCN (DGCNN)	ShapeNet	93.2
DHGCN (AdaptConv)	ShapeNet	93.2
DHGCN (DGCNN)	ModelNet40	93.0
DHGCN (AdaptConv)	ModelNet40	93.3

Shape Part Segmentation

• Multi-task Learning

Methods	Class mIoU	Instance mIoU
DGCNN	82.3	85.2
DHGCN (DGCNN)	82.8 (+0.5)	85.5 (+0.3)
PAConv	84.2	86.0
DHGCN (PAConv)	84.5 (+0.3)	86.1 (+0.1)

• Unsupervised Representation Learning

Methods	Class mIoU	Instance mIoU
DHGCN (PAConv)	82.9	84.9

• Visual Result (example)

Scene Segmentation

Methods	Accuracy	Instance mIoU
DGCNN	84.1	56.1
DHGCN (DGCNN)	86.4 (+2.3)	61.7 (+5.6)

• Visual Result (example)

Citation

If you find our work useful, please consider citing:

@inproceedings{jiang2024dhgcn,
  title={DHGCN: Dynamic Hop Graph Convolution Network for Self-Supervised Point Cloud Learning},
  author={Jiang, Jincen and Zhao, Lizhi and Lu, Xuequan and Hu, Wei and Razzak, Imran and Wang, Meili},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={11},
  pages={12883-12891},
  year={2024}
}