PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

Guocheng Qian; Yuchen Li; Houwen Peng; Jinjie Mai; Hasan Abed Al Kader Hammoud; Mohamed Elhoseiny; Bernard Ghanem

Source-linked AI summary

PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, Bernard Ghanem

arXiv:2206.04670v2Jun 9, 2022cs.CVcs.AI

TL;DR

PointNet++ appeared to lag behind newer point-cloud networks, but the roles of training strategies and model scaling were underexplored. This paper systematically modernizes PointNet++ and introduces PointNeXt, which surpasses state-of-the-art methods across studied tasks while maintaining faster inference.

Problem
PointNet++ was outperformed by newer point-cloud networks, leaving the effects of training strategies and effective model scaling insufficiently examined.
Method
The paper systematically studies training, receptive-field, and model-scaling strategies, then modernizes PointNet++ with an inverted residual MLP design to create PointNeXt.
Results
PointNeXt surpasses state-of-the-art methods across studied classification, semantic-segmentation, and part-segmentation tasks while achieving faster inference.
Takeaways & Limitations
Improved training and scaling strategies substantially raise PointNet++ performance and provide a scalable, high-throughput successor for point-cloud analysis.
Takeaways & Limitations
PointNeXt-XL reaches 44M parameters but the study does not scale further because point-cloud datasets are smaller than large image datasets.

Abstract

from arXiv · show

PointNet++ is one of the most influential neural architectures for point cloud understanding. Although the accuracy of PointNet++ has been largely surpassed by recent networks such as PointMLP and Point Transformer, we find that a large portion of the performance gain is due to improved training strategies, i.e. data augmentation and optimization techniques, and increased model sizes rather than architectural innovations. Thus, the full potential of PointNet++ has yet to be explored. In this work, we revisit the classical PointNet++ through a systematic study of model training and scaling strategies, and offer two major contributions. First, we propose a set of improved training strategies that significantly improve PointNet++ performance. For example, we show that, without any change in architecture, the overall accuracy (OA) of PointNet++ on ScanObjectNN object classification can be raised from 77.9% to 86.1%, even outperforming state-of-the-art PointMLP. Second, we introduce an inverted residual bottleneck design and separable MLPs into PointNet++ to enable efficient and effective model scaling and propose PointNeXt, the next version of PointNets. PointNeXt can be flexibly scaled up and outperforms state-of-the-art methods on both 3D classification and segmentation tasks. For classification, PointNeXt reaches an overall accuracy of 87.7 on ScanObjectNN, surpassing PointMLP by 2.3%, while being 10x faster in inference. For semantic segmentation, PointNeXt establishes a new state-of-the-art performance with 74.9% mean IoU on S3DIS (6-fold cross-validation), being superior to the recent Point Transformer. The code and models are available at https://github.com/guochengqian/pointnext.

1 Introduction

The introduction argues that training strategies and model scaling substantially improve PointNet++ performance, with gains often attributed to architectural changes. It motivates PointNeXt as a scalable successor built on these findings.

Motivation: Training strategies, including data augmentation and optimization, play an important role in point cloud network performance.The study examines benchmarks including ScanObjectNN for object classification and S3DIS for semantic segmentation.
Training strategies: 5.9% mIoU is gained on S3DIS by randomly dropping colors during PointNet++ training.Label smoothing improves PointNet++ by 1.3% OA on ScanObjectNN.
Model scaling: PointNeXt is proposed as a scalable next version of PointNets after observing that prevailing models use more parameters than original PointNets.The introduction frames larger models as potentially enabling richer representations and better performance, while identifying naive expansion of PointNet++ as insufficient.
Key contributions: +8.2% OA on ScanObjectNN and +13.6% mIoU on S3DIS are achieved by PointNet++ using improved training strategies alone.The paper presents this as a systematic study whose strategies are general and can be applied to other methods.

2 Preliminary: A Review of PointNet++

PointNeXt is built on PointNet++, a U-Net-like point-cloud architecture with hierarchical encoding and feature-propagation decoding. The paper uses single-scale PointNet++ with one set-abstraction block per stage as its baseline.

Architecture: PointNet++ uses a U-Net-like architecture with an encoder of set-abstraction blocks and a decoder of feature-propagation blocks.The encoder hierarchically abstracts point-cloud features, while the decoder gradually interpolates the abstracted features.
Architecture: Each set-abstraction block subsamples incoming points and groups neighboring points for local feature processing.
Feature computation: Neighbor features are aggregated with a reduction layer, such as max-pooling, and transformed by shared MLPs using features and relative coordinates.The reduction layer aggregates features from neighbors queried for each point.
Baseline: The default PointNet++ baseline uses single-scale grouping with one set-abstraction block per stage.This configuration matches the architecture used in the original PointNet++ paper.

3 Methodology: From PointNet++ to PointNeXt

PointNeXt modernizes PointNet++ through systematic improvements to training, receptive-field scaling, and model scaling. Its architecture adds InvResMLP blocks and macro-architectural changes while retaining PointNet++’s core Set Abstraction and Feature Propagation structure.

Methodology focus: PointNeXt’s methodology targets training modernization and architectural modernization, focusing on data augmentation, optimization, receptive-field scaling, and model scaling.These aspects were identified as important but under-explored influences on PointNet++ performance.
Training modernization: The training study starts from PointNet++ with its original augmentations and optimization, removes augmentations individually, and systematically evaluates techniques from representative point-cloud networks.The study includes point resampling and loading entire scenes as input among other data-scaling strategies.
Training modernization: Modern optimization replaces PointNet++’s CrossEntropy, Adam, Step Decay, and fixed hyperparameters with techniques including label smoothing, AdamW, and Cosine Decay.The paper identifies loss functions, optimizers, learning-rate schedulers, and hyperparameters as vital to network performance.
Receptive-field scaling: Receptive-field scaling studies larger neighborhood-query radii, while relative position normalization divides relative positions by the query radius to ease optimization and reduce variance across stages.PointNet++ already uses a hierarchical architecture, so the subsection focuses primarily on increasing the query radius.
Model scaling: PointNeXt scales PointNet++ with per-stage InvResMLP blocks, residual connections, separable MLPs, a unified encoder, a symmetric decoder, and a stem MLP.InvResMLP blocks support efficient scaling through residual learning and separated pointwise feature extraction.

4 Experiments

Experiments across classification and segmentation benchmarks show that improved training substantially strengthens PointNet++, while PointNeXt’s scaling and architectural changes deliver further gains and efficiency improvements.

Semantic segmentation: 57.2% validation mIoU versus 53.5% improves PointNet++ on ScanNet when adopting the improved training strategies.The comparison follows public training, validation, and test splits and uses Stratified Transformer results as the PointNet++ reference.
Scaling studies: Over 4× faster performance is achieved by the width-scaled PointNeXt model that outperforms CurveNet on ShapeNetPart, while naive scaling strategies are slower and less effective.The default-width model is comparable to CurveNet, and increasing width from C = 32 to 64 improves the comparison before the larger-width result.
Ablation studies: 2.5% OA is gained from point resampling in ScanObjectNN, while data scaling also improves segmentation when the entire scene replaces block or sphere subsampling.These observations come from additive studies of sequentially applied training and scaling strategies using original PointNet++ as the baseline.
Training strategies: 86.1% OA versus 77.9% raises PointNet++ performance on ScanObjectNN and surpasses PointMLP by 0.7% through stronger augmentation and modern optimization.The same strategies boost S3DIS area 5 and 6-fold mIoU by 11.7 and 13.6 absolute percentage points, respectively.
Architectural ablations: 3.9% mIoU is added by separable MLPs while speeding the network 3 times, whereas removing residual connections drops mIoU from 70.5% to 64.0%.The ablation uses PointNeXt-XL on S3DIS area 5 and finds residual connections are the most essential architectural change.

5 Related Work

Related work spans point-based architectures, improved training strategies, and model scaling. PointNet introduced permutation-invariant pointwise feature extraction, PointNet++ added local geometric modeling, and subsequent studies showed that training and scaling substantially affect point-cloud network performance.

Point-based methods: Point-based methods process unstructured point clouds directly; PointNet uses shared MLPs for permutation-invariant pointwise feature extraction, while PointNet++ captures local geometric structures.These methods are contrasted with voxel-based and multi-view approaches.
Training strategies: SimpleView showed that training strategies strongly affect point-cloud network performance, extending earlier studies from image classification.SimpleView adopts the same training strategies as DGCNN.
Training strategies: The paper systematically quantifies data augmentation and optimization techniques and proposes improved strategies for PointNet++ and other representative point-based methods.The proposed strategies are designed to boost performance without being limited to PointNet++.
Model scaling: Model scaling can significantly improve network performance, motivating effective and efficient strategies for scaling PointNet++.PointNet++ uses less than 2M parameters, whereas KPConv and PointMLP use 15M and 13M, respectively.

6 Conclusion and Discussion

The paper concludes that improved training and scaling strategies can raise PointNet++ performance beyond the current state of the art and be applied to other representative works. It also notes that PointNeXt-XL remains smaller than large image-classification models because point clouds are smaller-scale.

Conclusion: Improved training and scaling strategies increase PointNet++ performance beyond the current state of the art.The authors quantify the effects of widely used data augmentation and optimization techniques and propose improved training strategies applicable to PointNet++ and other representative works.
Limitation: 44M parameters make PointNeXt-XL smaller than Swin-S (50M), ConNeXt-S (50M), and ViT-B (87M).The paper describes PointNeXt-XL as one of the largest representative point-based networks, while noting that it remains below these small image-classification models.
Limitation: PointNeXt-XL remains far smaller than Swin-L (197M), ConvNeXt-XL (350M), and ViT-L (305M).The authors did not scale the model further, mainly because point clouds are smaller-scale than image-classification settings.

A Detailed Description for Manuscript Tab. 7

The passage compares naive width, depth, and compound scaling strategies for PointNet++ against PointNeXt-XL throughput and an improved-training PointNet++ model.

Scaling strategies: Naive width scaling expands PointNet++ channels from 32 to 256 to match PointNeXt-XL throughput.The passage identifies PointNeXt-XL as the baseline model whose throughput is matched.
Scaling strategies: Naive depth scaling appends SA blocks with B = (3, 6, 3, 3), matching PointNeXt-XL.This configuration increases PointNet++ depth using additional set-abstraction blocks.
Scaling strategies: Naive compound scaling doubles the width of the depth-scaled model to C = 64, matching PointNeXt-XL.The compound-scaled model uses the same target width as PointNeXt-XL.

B Training Strategies Comparison

This section summarizes the training strategies used by representative point-based methods across four point-cloud benchmarks. It organizes the comparison by dataset and corresponding table.

The comparison covers training strategies for DGCNN, KPConv, PointMLP, Point Transformer, Stratified Transformer, PointNet++, and PointNeXt.
The methods are compared on S3DIS, ScanObjectNN, ScanNet, and ShapeNetPart.
The results are organized in Tables I–IV, respectively corresponding to S3DIS, ScanObjectNN, ScanNet, and ShapeNetPart.

C Qualitative Results

The paper presents qualitative comparisons showing that PointNeXt predictions are closer to ground truth than PointNet++ on S3DIS and ShapeNetPart. On S3DIS, PointNeXt also segments difficult classes such as doors and clutter.

Qualitative comparisons: Qualitative results compare PointNeXt-XL on S3DIS and PointNeXt-S (C = 160) on ShapeNetPart against PointNet++.The figures include PointNet++ trained with its original training strategies.
Qualitative comparisons: PointNeXt produces predictions closer to ground truth than PointNet++ on both S3DIS and ShapeNetPart.
S3DIS: On S3DIS, PointNeXt segments difficult classes including doors and clutter.Doors are highlighted in the first, third, and fourth rows of the qualitative results.

D Classification Architecture

The classification architecture shares the segmentation encoder, then applies global max-pooling to obtain a global shape representation. Because classification inputs are typically small, points are downsampled by only 2× at each stage.

Classification Architecture: The classification network reuses the segmentation architecture’s encoder and feeds its output features into global max-pooling for shape-level representation.This pooled global representation is used for classification.

E Societal Impact

The authors identify no immediate negative societal impact, but acknowledge modest additional computing and environmental costs from discovering improved training and scaling strategies. They argue these strategies may ultimately broaden and improve computer vision research by shifting attention beyond architectural changes.

The authors report no immediate negative societal impact from the work.
Discovering the improved strategies may consume additional computing resources and affect the environment.The authors characterize this impact as limited.
The strategies may encourage researchers to focus beyond architectural changes, making computer vision research more diverse and generally better over time.