Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework

Xu Ma; Can Qin; Haoxuan You; Haoxi Ran; Yun Fu

Source-linked AI summary

Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework

Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, Yun Fu

arXiv:2202.07123v2Feb 15, 2022cs.CVcs.AI

TL;DR

Point cloud analysis must handle unordered, irregular data, while elaborate local geometric extractors impose efficiency costs and face performance saturation. PointMLP addresses this with residual MLPs and a lightweight geometric affine module, achieving competitive accuracy without sophisticated extractors. The method reports state-of-the-art results across benchmarks, including 94.5% on ModelNet40 and a 3.3% ScanObjectNN improvement over related work.

Problem
Sophisticated local geometric extractors improve point-cloud representations but introduce unfavorable computational and memory-access costs as performance saturates.
Method
PointMLP uses residual feed-forward MLPs without delicate local feature extractors and adds a lightweight geometric affine module for local feature transformation.
Results
PointMLP achieves state-of-the-art performance across benchmarks, including 94.5% accuracy on ModelNet40 and a 3.3% accuracy advantage on ScanObjectNN.
Takeaways & Limitations
Strong results without sophisticated local extractors support rethinking whether elaborate local geometry is essential for point-cloud analysis.
Takeaways & Limitations
Model complexity and speed are not fully captured by FLOPs or parameter counts because memory access cost and parallelism also matter.

Abstract

from arXiv · show

Point cloud analysis is challenging due to irregularity and unordered data structure. To capture the 3D geometries, prior works mainly rely on exploring sophisticated local geometric extractors using convolution, graph, or attention mechanisms. These methods, however, incur unfavorable latency during inference, and the performance saturates over the past few years. In this paper, we present a novel perspective on this task. We notice that detailed local geometrical information probably is not the key to point cloud analysis -- we introduce a pure residual MLP network, called PointMLP, which integrates no sophisticated local geometrical extractors but still performs very competitively. Equipped with a proposed lightweight geometric affine module, PointMLP delivers the new state-of-the-art on multiple datasets. On the real-world ScanObjectNN dataset, our method even surpasses the prior best method by 3.3% accuracy. We emphasize that PointMLP achieves this strong performance without any sophisticated operations, hence leading to a superior inference speed. Compared to most recent CurveNet, PointMLP trains 2x faster, tests 7x faster, and is more accurate on ModelNet40 benchmark. We hope our PointMLP may help the community towards a better understanding of point cloud analysis. The code is available at https://github.com/ma-xu/pointMLP-pytorch.

1 INTRODUCTION

PointMLP questions whether elaborate local geometric extractors are necessary for point cloud analysis. It uses residual MLPs plus a lightweight geometric affine module to achieve strong accuracy with improved efficiency.

Point clouds are unordered and irregular, making direct application of standard image-processing methods infeasible.
Sophisticated convolution-, graph-, and attention-based local extractors increase computational and memory-access costs while performance has begun to saturate.
PointMLP uses residual feed-forward MLPs without delicate local feature exploration, supplemented by a lightweight geometric affine module.
PointMLP is presented as more accurate and faster than methods based on sophisticated local geometric extractors.
94.5% classification accuracy establishes PointMLP as state of the art on ModelNet40, while ScanObjectNN accuracy exceeds related work by 3.3%.

2 RELATED WORK

Related point-cloud methods either project data into voxels or images or process unordered point sets directly. Direct methods increasingly emphasize local geometry through convolution, graph, and attention mechanisms.

Projection-based methods translate point clouds into voxel or image representations, gaining processing speed but losing detailed information.
PointNet directly processes unordered point sets with shared MLPs, while PointNet++ hierarchically learns local geometric structures.
Recent local-representation methods mainly use convolution-, graph-, or attention-based mechanisms.
Point-cloud network development has parallels with image-processing architectures, including the progression toward deeper neural networks.
Representative methods are systematically compared by whether they are expandable in depth and by their principal operator.

3 DEEP RESIDUAL MLP FOR POINT CLOUD

PointMLP revisits point-based point-cloud analysis by replacing elaborate local geometric extractors with residual MLPs and a lightweight geometric affine module. Its staged design hierarchically aggregates local features while targeting lower computational cost and competitive accuracy.

Method: PointMLP learns point-cloud representations with a feed-forward residual MLP network that hierarchically aggregates local features without delicate local geometric extractors.Its key operation applies shared pre-aggregation and post-aggregation residual point MLP blocks.
Point-based methods: Point-based methods directly process unordered point sets, using sampled neighborhoods and aggregation to capture local structures and progressively enlarge receptive fields.The PointNet++ pipeline samples points, gathers K neighbors, applies a local feature extractor, and aggregates the resulting features.
Architecture: PointMLP repeats stages that transform local points, extract features before and after aggregation, and progressively enlarge the receptive field.The geometric affine module precedes local feature extraction within each stage.
Motivation: Convolution-, graph-, and attention-based local extractors improve regional representations but increase computational complexity and inference latency.The paper highlights prohibitive inference costs associated with delicate extractors and notes that memory access cost is often overlooked.
Geometric affine module: The geometric affine module normalizes grouped local features with learnable affine parameters while maintaining their original geometric properties.It uses a small epsilon for numerical stability and addresses reduced accuracy and stability observed in deeper simple MLP structures.
Results: PointMLP achieves 94.5% overall accuracy on ModelNet40 and outperforms CurveNet by 0.3 percentage points with only 1k points.The table also reports classification accuracy and speed for open-source methods under specified GPU and CPU test conditions.

4 EXPERIMENTS

Experiments evaluate PointMLP across synthetic and real-world classification, ablations, efficiency, and segmentation settings. PointMLP achieves strong accuracy and speed, while the geometric affine module and residual connections improve performance, stability, or optimization.

Shape Classification on ModelNet40: PointMLP achieves 94.5% overall accuracy on ModelNet40, outperforming CurveNet by 0.3% with only 1k points.The comparison is reported without relying on a voting strategy.
Shape Classification on ModelNet40: 112 samples/second inference speed makes PointMLP faster than CurveNet’s 15 samples/second while maintaining higher accuracy.PointMLP-elite further reaches 176 samples/second with 0.68M parameters and 94.0% overall accuracy.
Shape Classification on ScanObjectNN: Four-run evaluation on the hardest ScanObjectNN variant reports PointMLP at 83.9±0.5 mAcc and 85.4±0.3 OA.The benchmark contains real-world objects with background, noise, and occlusions; results are reported as mean ± standard deviation.
Shape Classification on ScanObjectNN: 3.3% overall-accuracy improvement over PRANet is reported on ScanObjectNN, alongside a 4.8% class-mean-accuracy gain.PointMLP also achieves 85.4% overall accuracy, compared with MVTN’s 82.8%, without voting and with fewer training epochs.
Ablation Studies: The geometric affine module improves PointMLP variants by about 3% and increases stability, while component ablation reports a 3.9% improvement over the base architecture.Removing the Φpre function reduces overall accuracy by 2.6%, whereas combining all components reaches 85.4% OA.
Ablation Studies: Removing residual connections sharpens the loss landscape and reduces ModelNet40 performance to 88.1%, a 6% drop.Residual connections ease optimization and support training a deep network.

5 CONCLUSION

PointMLP argues that simple residual MLPs, combined with a lightweight geometric affine module, can provide strong point-cloud analysis without sophisticated local extractors.

PointMLP uses simple residual MLPs to represent local points because they are permutation-invariant and straightforward.
A lightweight geometric affine module boosts PointMLP's performance, while PointMLP-elite further reduces efficiency costs.
PointMLP outperforms related work on different benchmarks beyond simplicity and efficiency.

A POINTMLP DETAIL

The detailed architecture describes a lighter PointMLP-elite variant and reports a backbone replacement that improves part-segmentation performance.

PointMLP-elite reduces residual point MLP blocks and embedding dimension from 64 to 32, lowering overall model overhead.
A bottleneck structure reduces PointMLP-elite's parameters by four times compared with PointMLP.
Replacing the PointNet backbone with PointMLP improves part-segmentation performance from 85.1 to 86.1 Instance mIoU.
Figures 6 and 7 detail the PointMLP and PointMLP-elite classification architectures.

B.1 MODELNET40 AND SCANOBJECTNN

The ModelNet40 and ScanObjectNN experiments use specified training and speed-testing setups, with reported implementations and source-code references for comparison methods.

ModelNet40 models are trained for 300 epochs on one Tesla V100 GPU with batch size 32.
All models use synchronous SGD with Nesterov momentum 0.9, weight decay 0.0002, an initial learning rate of 0.1, and cosine annealing.
Reported speed is measured using open-source code on a Tesla V100-pcie GPU.
The source codes used for PointNet++, CurveNet, GBNet, GDANet, PointConv, and KPConv are listed in the footnote.

B.2 SHAPENETPART

The part-segmentation setup follows PointNet, samples 2048 points, rescales inputs, and reports single-scale results without multi-scale testing.

The part-segmentation setting follows PointNet, randomly sampling 2048 points for each sample.
Inputs are re-scaled to a range of [0.67, 1.5].
Results are reported using single-scale testing because multi-scale testing is not realizable in real-world applications.
Despite the authors' stated unfair comparison, the method achieves competitive performance.

C MORE DETAILED ABLATION STUDIES

The ablations examine skip connections, block placement, and the geometric affine module, showing that architectural choices affect PointMLP performance and applicability.

Skip connection: 92.7% classification accuracy is achieved on ModelNet40 after adding skip connections to the re-implemented PointNet++.Only two skip connections could be added without modifying PointNet++’s original architecture.
Pre-MLP block vs. Pos-MLP block: 85.4% vanilla PointMLP accuracy exceeds the 84.13% average obtained when pos-MLP blocks are removed and pre-MLP blocks are increased.The modified models achieved 83.87% and 84.39% across two training runs.
Pre-MLP block vs. Pos-MLP block: The ablation indicates that pos-MLP blocks benefit PointMLP, whereas simply adding more pre-MLP blocks does not help.The authors describe pos-MLP’s effect as weaker than those of other components.
Geometric Affine Module Applications: The geometric affine module is presented as essential to PointMLP and as a plug-and-play method whose overlap with local geometric extractors may limit applications.The module is integrated into PointNet++ and DGCNN for ModelNet40 experiments, but the supplied passage does not report their resulting improvements.

D POINTMLP DEPTH

PointMLP’s depth is determined by repeated pre-MLP and pos-MLP blocks across stages, together with embedding, channel-matching, and classifier layers. The reported depth configurations are 24, 40, and 56.

Layer-count formulation: PointMLP’s learnable-layer count includes repeated Φpre and Φpos blocks across stages, plus embedding, channel-matching, and classifier layers.Φpre contains three layers and Φpos contains two layers; the formulation ignores Batch Normalization and activation layers.