[Paper Review] RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds

Qingyong Hu, Bo Yang+6 more|arXiv (Cornell University)|Nov 25, 2019

3D Shape Modeling and Analysis68 references133 citations

TL;DR

RandLA-Net performs per-point semantic segmentation on large-scale 3D point clouds using random sampling plus a lightweight local feature aggregator, achieving state-of-the-art results on Semantic3D and SemanticKITTI while being significantly faster and more memory-efficient than prior methods.

ABSTRACT

We study the problem of efficient semantic segmentation for large-scale 3D point clouds. By relying on expensive sampling techniques or computationally heavy pre/post-processing steps, most existing approaches are only able to be trained and operate over small-scale point clouds. In this paper, we introduce RandLA-Net, an efficient and lightweight neural architecture to directly infer per-point semantics for large-scale point clouds. The key to our approach is to use random point sampling instead of more complex point selection approaches. Although remarkably computation and memory efficient, random sampling can discard key features by chance. To overcome this, we introduce a novel local feature aggregation module to progressively increase the receptive field for each 3D point, thereby effectively preserving geometric details. Extensive experiments show that our RandLA-Net can process 1 million points in a single pass with up to 200X faster than existing approaches. Moreover, our RandLA-Net clearly surpasses state-of-the-art approaches for semantic segmentation on two large-scale benchmarks Semantic3D and SemanticKITTI.

Motivation & Objective

• Motivate efficient semantic segmentation for large-scale, irregular 3D point clouds without heavy pre/post-processing.
• Demonstrate that random sampling can be effective when paired with a robust local feature aggregator.
• Propose LocSE-based local spatial encoding and attentive pooling to preserve geometry across downsampling.
• Show that RandLA-Net achieves substantial speedups and memory savings while maintaining or surpassing prior accuracy on benchmarks.

Proposed method

• Use random sampling to downsample large-scale point clouds in a single pass, avoiding expensive FPS/IDIS-based methods.
• Introduce a Local Spatial Encoding (LocSE) unit to explicitly embed relative neighborhood geometry.
• Apply attentive pooling to weigh and combine neighboring features adaptively.
• Stack LocSE and attentive pooling into a dilated residual block to progressively enlarge the receptive field.
• Construct the network with lightweight shared MLPs, avoiding graph construction or voxelization steps.
• Train end-to-end with Adam on fixed-point subsets (~1e5 points) and test on full clouds without pre/post-processing.

Experimental results

Research questions

• RQ1 Can random sampling enable real-time or near-real-time semantic segmentation of multi-million point clouds without heavy preprocessing?
• RQ2 How can local geometry and features be preserved when random sampling downscales data?
• RQ3 Does a dilated residual arrangement of LocSE and attentive pooling effectively enlarge the receptive field for large-scale point clouds?
• RQ4 What are the efficiency and accuracy trade-offs of RandLA-Net compared with state-of-the-art methods on Semantic3D and SemanticKITTI?

Key findings

• RandLA-Net processes up to 1 million points in a single pass and is up to 200× faster than existing approaches on large-scale point clouds.
• It directly processes large-scale point clouds (e.g., up to 1e6 points) without pre/post-processing, achieving state-of-the-art results on Semantic3D and SemanticKITTI.
• The network’s 0.04s per 1e5-point cloud demonstration and 22 FPS on Sequence 08 (SemanticKITTI) illustrate strong real-time capability.
• RandLA-Net achieves 77.4 mIoU on Semantic3D and 53.9 mIoU (with 50k input points) on SemanticKITTI, outperforming many baselines while using fewer parameters.
• Ablation shows removing LocSE or the attention module substantially degrades performance, confirming the effectiveness of the local geometry encoding and adaptive feature weighting.