[Paper Review] PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

Guocheng Qian, Yuchen Li+5 more|arXiv (Cornell University)|Jun 9, 2022

3D Surveying and Cultural Heritage412 citations

TL;DR

This paper re-evaluates PointNet++ with modern training and scaling strategies, introduces PointNeXt, and shows state-of-the-art results on several 3D tasks.

ABSTRACT

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.

Motivation & Objective

• Motivate a systematic study of training and scaling strategies for PointNet++ to uncover the network's full potential.
• Demonstrate that improved training alone can significantly boost PointNet++ performance on benchmarks.
• Propose architectural and scaling enhancements to create PointNeXt for scalable, efficient 3D point cloud understanding.
• Show that PointNeXt achieves superior accuracy and speed across classification and segmentation tasks.

Proposed method

• Systematic evaluation of data augmentations and optimization techniques on PointNet++ to identify effective training improvements.
• Incorporation of relative position normalization to stabilize and accelerate training.
• Architectural modernization via InvResMLP blocks with residual connections and separable MLPs for efficient scaling.
• Macro-architectural changes including stem MLP, symmetric decoder, and revised encoder depth to enable scalable PointNeXt variants.
• Definition of PointNeXt family (S, B, L, XL) with controlled width and depth scaling parameters.

Experimental results

Research questions

• RQ1 Can modern training strategies (data augmentation and optimization) recover and surpass PointNet++ performance without changing the architecture?
• RQ2 How can PointNet++ be scaled effectively and efficiently to compete with state-of-the-art point cloud networks?
• RQ3 What architectural changes (normalization, residual connections, separable MLPs) most improve performance and scalability?
• RQ4 Do the proposed training and architectural modifications translate into gains across classification and segmentation benchmarks?
• RQ5 How do different PointNeXt configurations perform in terms of accuracy and inference speed across tasks?

Key findings

• Improved training strategies alone raise PointNet++ OA on ScanObjectNN from 77.9% to 86.1%, surpassing several SOTA methods.
• On S3DIS segmentation (6-fold), training improvements raise mean IoU from 54.5% to 68.1%, outperforming several later architectures.
• PointNeXt-S/B/L/XL achieve strong gains, e.g., PointNeXt-XL reaches 74.9% mIoU on S3DIS 6-fold and 90.3% OA on ScanObjectNN, while being faster than some baselines.
• PointNeXt-L can exceed state-of-the-art Point Transformer in mIoU while maintaining faster inference; PointNeXt-XL provides the best reported results among the studied variants.
• In object classification on ScanObjectNN, PointNeXt-S achieves 87.7% OA, outperforming PointMLP and being 10x faster in inference.
• ShapeNetPart results with PointNeXt-S (C=64) reach competitive or superior ins. mIoU and cls. mIoU, with substantial throughput gains; very large width (C=160) yields 87.0% ins. mIoU.