Revisiting Point Cloud Shape Classification with a Simple and Effective Baseline

Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, Jia Deng

Introduction

Over the past few years, accuracy on 3D point cloud classification benchmarks has steadily climbed. The research community has generally attributed this progress to increasingly sophisticated "point-based" neural network architectures (like PointConv or RSCNN) that process 3D coordinates directly.

However, recent performance gains in 3D point cloud classification may be largely due to training protocols—such as data augmentation, loss functions, and evaluation schemes—rather than novel network architectures. By controlling for these factors, older models like PointNet++ remain highly competitive. Furthermore, a remarkably simple 2D projection-based baseline, dubbed SimpleView, can outperform state-of-the-art 3D point-based methods while using fewer parameters.

ModelNet40 accuracy over time, showing the impact of standardized protocols.

The core challenge addressed in this work is disentangling architecture improvements from protocol improvements. Are new architectures actually better, or are they just using better training pipelines? And are point-based methods inherently superior to older projection-based methods, as the community currently assumes?

Methodology

To answer these questions, the authors employ two main approaches: a rigorous protocol analysis and the introduction of a new baseline.

1. Rigorous Protocol Analysis

The authors standardize factors like data augmentation (e.g., random scaling and translation), loss functions (e.g., label smoothing), and model selection. Crucially, they avoid the flawed practice of using test-set feedback to pick the best training epoch. They then re-evaluate older architectures (PointNet, PointNet++) and newer ones (DGCNN, RSCNN) under these unified, fair protocols.

2. The SimpleView Baseline

To challenge the assumption that 3D point-based methods are strictly necessary, the authors introduce a stripped-down projection-based baseline.

Architecture of the SimpleView baseline.

Instead of processing 3D points directly, SimpleView projects the point cloud onto six orthogonal planes to create sparse 2D depth images. These images are fed into a lightweight CNN (a scaled-down ResNet18), and the resulting features are concatenated for classification. Notably, SimpleView avoids the complex rendering, view selection, or ImageNet pretraining used by prior projection-based methods.

Datasets

The models were evaluated on two primary datasets:

  • ModelNet40: The most widely adopted benchmark for 3D point cloud classification, consisting of synthetic, clean, and axis-aligned CAD models across 40 categories.
  • ScanObjectNN: A newer, real-world dataset containing noisy point clouds (with background points, occlusions, and holes) that are not axis-aligned. It is used to test robustness and real-world applicability.

Key Insights and Results

  • Protocols Obscure Architectural Gains: Changing the training and evaluation protocol can cause massive swings in performance. For example, simply applying a newer protocol to the older PointNet++ architecture boosts its accuracy from 90.0% to 93.3%, making it highly competitive with the latest models.
  • Test-Set Feedback is a Hidden Crutch: Some recent "state-of-the-art" methods achieved their high numbers by evaluating on the test set after every epoch and reporting the best score. This practice leaks test data into model selection and inflates perceived performance.
  • Projection Methods are Highly Underrated: The community largely shifted to point-based methods under the assumption that projecting 3D data to 2D loses critical geometric information. SimpleView proves this wrong, showing that a basic multi-view 2D projection can capture enough geometry to beat complex 3D convolutions.

Performance Metrics

  • On ModelNet40, under a strictly controlled and fair protocol, SimpleView achieves 93.0% accuracy, outperforming PointNet++ (92.7%), DGCNN (91.9%), and RSCNN (91.7%).
  • SimpleView is highly efficient, requiring only 0.8M parameters (less than half of PointNet++'s 1.7M) while maintaining competitive inference speeds (5.0 ms).
  • On the real-world ScanObjectNN dataset, SimpleView achieves state-of-the-art accuracy of 79.5%, beating prior methods like DGCNN (78.1%) and PointNet++ (77.9%).
  • SimpleView also demonstrated superior cross-dataset generalization (training on ScanObjectNN and testing on ModelNet40, and vice versa) compared to point-based architectures.

Conclusion and Future Work

While the current analysis is limited to shape classification, the authors suggest expanding this rigorous protocol analysis and the SimpleView baseline to more complex 3D tasks, such as scene and part segmentation.

These findings are not meant to discourage research into point-based methods. Instead, they serve as a wake-up call for the community to use controlled baselines. It remains entirely possible that future point-based innovations will genuinely surpass projection-based methods once protocols are properly controlled and evaluated fairly.

Summary updated

Abstract

Processing point cloud data is an important component of many real-world systems. As such, a wide variety of point-based approaches have been proposed, reporting steady benchmark improvements over ...
BibTeX
@inproceedings{goyal2021revisiting,
  author = {Ankit Goyal and Hei Law and Bowei Liu and Alejandro Newell and Jia Deng},
  title = {{Revisiting Point Cloud Shape Classification with a Simple and Effective Baseline}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=TVYsbpBA5t},
}
Figure