Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Point-GR: Graph Residual Point Cloud Network for 3D Object Classification and Segmentation

Md Meraz, Md Afzal Ansari, Mohammed Javed, Pavan Chakraborty

TL;DR

This work tackles 3D shape analysis on unordered point clouds by addressing permutation invariance and efficient feature extraction. It introduces Point-GR, a graph residual network that first transforms points into higher-dimensional representations via a Point-GR Transformation and then learns multi-scale features with a Point-GR Feature Learning Network. Key contributions include a k-NN graph construction with edge-aware features, residual learning to improve gradient flow, and a lightweight multi-scale architecture that achieves strong indoor-scene segmentation results on S3DIS (mean IoU 73.47%), while remaining competitive on ModelNet-40 and ShapeNet-Part. The proposed approach supports robust 3D perception for robotics and autonomous systems with reduced parameter counts and practical scalability.

Abstract

In recent years, the challenge of 3D shape analysis within point cloud data has gathered significant attention in computer vision. Addressing the complexities of effective 3D information representation and meaningful feature extraction for classification tasks remains crucial. This paper presents Point-GR, a novel deep learning architecture designed explicitly to transform unordered raw point clouds into higher dimensions while preserving local geometric features. It introduces residual-based learning within the network to mitigate the point permutation issues in point cloud data. The proposed Point-GR network significantly reduced the number of network parameters in Classification and Part-Segmentation compared to baseline graph-based networks. Notably, the Point-GR model achieves a state-of-the-art scene segmentation mean IoU of 73.47% on the S3DIS benchmark dataset, showcasing its effectiveness. Furthermore, the model shows competitive results in Classification and Part-Segmentation tasks.

Point-GR: Graph Residual Point Cloud Network for 3D Object Classification and Segmentation

TL;DR

This work tackles 3D shape analysis on unordered point clouds by addressing permutation invariance and efficient feature extraction. It introduces Point-GR, a graph residual network that first transforms points into higher-dimensional representations via a Point-GR Transformation and then learns multi-scale features with a Point-GR Feature Learning Network. Key contributions include a k-NN graph construction with edge-aware features, residual learning to improve gradient flow, and a lightweight multi-scale architecture that achieves strong indoor-scene segmentation results on S3DIS (mean IoU 73.47%), while remaining competitive on ModelNet-40 and ShapeNet-Part. The proposed approach supports robust 3D perception for robotics and autonomous systems with reduced parameter counts and practical scalability.

Abstract

In recent years, the challenge of 3D shape analysis within point cloud data has gathered significant attention in computer vision. Addressing the complexities of effective 3D information representation and meaningful feature extraction for classification tasks remains crucial. This paper presents Point-GR, a novel deep learning architecture designed explicitly to transform unordered raw point clouds into higher dimensions while preserving local geometric features. It introduces residual-based learning within the network to mitigate the point permutation issues in point cloud data. The proposed Point-GR network significantly reduced the number of network parameters in Classification and Part-Segmentation compared to baseline graph-based networks. Notably, the Point-GR model achieves a state-of-the-art scene segmentation mean IoU of 73.47% on the S3DIS benchmark dataset, showcasing its effectiveness. Furthermore, the model shows competitive results in Classification and Part-Segmentation tasks.

Paper Structure

This paper contains 13 sections, 3 equations, 4 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Proposed Methodology
  3. Point-GR Transformation
  4. Graph Construction
  5. Point Residual Embedding
  6. Point-GR Feature Learning
  7. Point-GR Classification and Segmentation Architecture
  8. Experimental Results
  9. 3D Object Classification
  10. Part-Segmentation
  11. Indoor Scene Segmentation
  12. Ablation Study
  13. Conclusion

Figures (4)

  • Figure 1: The figure shows the Point-GR Transformation network. It consists of Graph Construction and Point Residual Embedding blocks.
  • Figure 2: Shows the Point-GR Features Learning Network(F.L.N). It consists of Multi-Scale Graph Construction and Point Embedding block.
  • Figure 3: Deep learning based architecture of (a) The extended Point-GR architecture for Classification Network which includes transformation networks, and it generates $m$-class scores for point cloud input. and (b) The extended Point-GR architecture for Part-Segmentation which includes transformation networks, and it generates $N\times c$ segmentation scores for point cloud data.
  • Figure 4: The visual representation shows how the network segments each shape into different parts based on the point cloud input using the open-source Meshlab visualizer. The figure uses a grid layout with two rows: the top row shows the ground truth with multi-colored point clouds, and the bottom row shows the prediction with variations in point cloud colors.