[Paper Review] ASSANet: An Anisotropic Separable Set Abstraction for Efficient Point Cloud Representation Learning

Guocheng Qian, Hasan Abed Al Kader Hammoud+3 more|arXiv (Cornell University)|Oct 20, 2021

3D Shape Modeling and AnalysisEngineering48 references40 citations

TL;DR

ASSANet replaces PointNet++'s SA with Anisotropic Separable SA (ASSA), achieving faster inference and higher accuracy on classification, segmentation, and part segmentation benchmarks.

ABSTRACT

Access to 3D point cloud representations has been widely facilitated by LiDAR sensors embedded in various mobile devices. This has led to an emerging need for fast and accurate point cloud processing techniques. In this paper, we revisit and dive deeper into PointNet++, one of the most influential yet under-explored networks, and develop faster and more accurate variants of the model. We first present a novel Separable Set Abstraction (SA) module that disentangles the vanilla SA module used in PointNet++ into two separate learning stages: (1) learning channel correlation and (2) learning spatial correlation. The Separable SA module is significantly faster than the vanilla version, yet it achieves comparable performance. We then introduce a new Anisotropic Reduction function into our Separable SA module and propose an Anisotropic Separable SA (ASSA) module that substantially increases the network's accuracy. We later replace the vanilla SA modules in PointNet++ with the proposed ASSA module, and denote the modified network as ASSANet. Extensive experiments on point cloud classification, semantic segmentation, and part segmentation show that ASSANet outperforms PointNet++ and other methods, achieving much higher accuracy and faster speeds. In particular, ASSANet outperforms PointNet++ by $7.4$ mIoU on S3DIS Area 5, while maintaining $1.6 imes $ faster inference speed on a single NVIDIA 2080Ti GPU. Our scaled ASSANet variant achieves $66.8$ mIoU and outperforms KPConv, while being more than $54 imes$ faster.

Motivation & Objective

• Motivate efficient point cloud processing for mobile and embedded devices.
• Investigate bottlenecks in PointNet++ SA and propose faster alternatives.
• Develop Separable SA and Anisotropic Reduction to improve accuracy without increasing latency.
• Replace PointNet++ SA with ASSA to form the ASSANet architecture.
• Demonstrate scalability via width and depth scaling of ASSANet.

Proposed method

• Introduce PreConv SA to apply MLPs on point features before grouping.
• Develop Separable SA that splits learning into pre-grouping and post-grouping MLPs with a residual connection.
• Propose Anisotropic Reduction that uses per-neighbor relative positions to scale features during reduction.
• Integrate ASSA into PointNet++ to form ASSANet and evaluate on classification, semantic segmentation, and part segmentation.
• Scale ASSANet by width (C) and depth (D) to create larger variants (e.g., ASSANet (L)).
• Provide comparative results against PointNet++, PVCNN, KPConv, PosPool, DeepGCN, and others.

Experimental results

Research questions

• RQ1 Can the SA bottlenecks in PointNet++ be mitigated without sacrificing accuracy?
• RQ2 Does separable processing and anisotropic reduction improve both speed and accuracy for neighborhood feature aggregation?
• RQ3 How does ASSANet perform on standard 3D tasks (classification, semantic segmentation, part segmentation) compared to state-of-the-art methods?
• RQ4 What is the impact of width and depth scaling on the accuracy-speed trade-off for ASSANet?

Key findings

• ASSANet achieves higher mIoU than PointNet++ on S3DIS Area-5 by 7.4 points while being 1.6x faster.
• ASSANet (L) reaches 66.8 mIoU on S3DIS Area-5, outperforming several methods in accuracy and being substantially faster (e.g., 294x faster than DeepGCN; 54.6x faster than KPConv).
• On ModelNet40 classification, ASSANet attains 92.4% OA (+1.7) with 2.1x speed, and ASSANet (L) reaches 92.9% OA (+2.2) with 1.3x speed.
• On ShapeNetPart, ASSANet and ASSANet (L) outperform PointNet++ and are notably faster; ASSANet (L) is ~7.8x faster than KPConv.
• Ablations show PreConv SA, Separable SA, and especially ASSA improve speed and accuracy versus vanilla SA, with Anisotropic Reduction providing the best accuracy gains.
• Scaling analyses reveal width and depth increases improve accuracy with manageable speed changes, with diminishing returns at higher widths.