Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Local Neighborhood Features for 3D Classification

Shivanand Venkanna Sheshappanavar, Chandra Kambhamettu

TL;DR

The paper addresses how to exploit precomputed neighborhood features to improve 3D point cloud classification. It introduces radius $r$-normalized distance and radius-normalized directional vectors as additional local features within PointNeXt, along with a checkpoint weight-averaging strategy for inference. Evaluations on ModelNet40, ScanObjectNN, and 3DGrocery100 show consistent improvements, particularly on real-world datasets, with modest computational overhead. Ablation analyses confirm the benefits of the proposed features and the efficacy of weight averaging, suggesting these techniques are broadly applicable to future 3D classifiers.

Abstract

With advances in deep learning model training strategies, the training of Point cloud classification methods is significantly improving. For example, PointNeXt, which adopts prominent training techniques and InvResNet layers into PointNet++, achieves over 7% improvement on the real-world ScanObjectNN dataset. However, most of these models use point coordinates features of neighborhood points mapped to higher dimensional space while ignoring the neighborhood point features computed before feeding to the network layers. In this paper, we revisit the PointNeXt model to study the usage and benefit of such neighborhood point features. We train and evaluate PointNeXt on ModelNet40 (synthetic), ScanObjectNN (real-world), and a recent large-scale, real-world grocery dataset, i.e., 3DGrocery100. In addition, we provide an additional inference strategy of weight averaging the top two checkpoints of PointNeXt to improve classification accuracy. Together with the abovementioned ideas, we gain 0.5%, 1%, 4.8%, 3.4%, and 1.6% overall accuracy on the PointNeXt model with real-world datasets, ScanObjectNN (hardest variant), 3DGrocery100's Apple10, Fruits, Vegetables, and Packages subsets, respectively. We also achieve a comparable 0.2% accuracy gain on ModelNet40.

Local Neighborhood Features for 3D Classification

TL;DR

The paper addresses how to exploit precomputed neighborhood features to improve 3D point cloud classification. It introduces radius -normalized distance and radius-normalized directional vectors as additional local features within PointNeXt, along with a checkpoint weight-averaging strategy for inference. Evaluations on ModelNet40, ScanObjectNN, and 3DGrocery100 show consistent improvements, particularly on real-world datasets, with modest computational overhead. Ablation analyses confirm the benefits of the proposed features and the efficacy of weight averaging, suggesting these techniques are broadly applicable to future 3D classifiers.

Abstract

With advances in deep learning model training strategies, the training of Point cloud classification methods is significantly improving. For example, PointNeXt, which adopts prominent training techniques and InvResNet layers into PointNet++, achieves over 7% improvement on the real-world ScanObjectNN dataset. However, most of these models use point coordinates features of neighborhood points mapped to higher dimensional space while ignoring the neighborhood point features computed before feeding to the network layers. In this paper, we revisit the PointNeXt model to study the usage and benefit of such neighborhood point features. We train and evaluate PointNeXt on ModelNet40 (synthetic), ScanObjectNN (real-world), and a recent large-scale, real-world grocery dataset, i.e., 3DGrocery100. In addition, we provide an additional inference strategy of weight averaging the top two checkpoints of PointNeXt to improve classification accuracy. Together with the abovementioned ideas, we gain 0.5%, 1%, 4.8%, 3.4%, and 1.6% overall accuracy on the PointNeXt model with real-world datasets, ScanObjectNN (hardest variant), 3DGrocery100's Apple10, Fruits, Vegetables, and Packages subsets, respectively. We also achieve a comparable 0.2% accuracy gain on ModelNet40.
Paper Structure (13 sections, 2 equations, 3 figures, 6 tables)

This paper contains 13 sections, 2 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Method
  4. Neighborhood point distance
  5. Directional Vectors as neighborhood features
  6. Weights-averaging of checkpoints
  7. Experiments
  8. Experimental Results
  9. Ablation Study
  10. Distance vs. $r$-normalized distance
  11. Weight averaging (same training session)
  12. Cost of additional neighborhood features
  13. Conclusion

Figures (3)

  • Figure 1: Airplane: Green points - input point cloud, Red point - anchor point for a neighborhood obtained from the farthest point sampling step, Blue points - points queried from the neighborhood of anchor point, Red Vectors - Directional Vectors from the anchor point to the neighbor points. The magnitude of each directional vector represents the distance of the neighbor point to its anchor point. The direction of these vectors is from the anchor point to the neighbor point.
  • Figure 2: Left: Input point cloud (green points), Farthest Point Sampled points (red points), Right: Directional vectors at each of the red points.
  • Figure 3: Few more visual examples of 3D objects with directional vectors after the grouping stage. From Left to Right: bottle, lamp, airplane, and chair.