[Paper Review] Projected-point-based Segmentation: A New Paradigm for LiDAR Point Cloud Segmentation

Shijie Li, Yun Liu+1 more|arXiv (Cornell University)|Aug 10, 2020

Remote Sensing and LiDAR Applications23 references4 citations

TL;DR

This paper proposes a new paradigm—projected-point-based segmentation—that adapts point-based methods for LiDAR point clouds in outdoor environments by leveraging the inherent ordering of LiDAR scans. By performing computations on projected images using only pointwise convolutions and matrix multiplication, the method achieves higher accuracy and efficiency than prior point-based approaches on the SemanticKITTI benchmark, even with a simple architecture.

ABSTRACT

Most point-based semantic segmentation methods are designed for indoor scenarios, but many applications such as autonomous driving vehicles require accurate segmentation for outdoor scenarios. For this goal, light detection and ranging (LiDAR) sensors are often used to collect outdoor environmental data. The problem is that directly applying previous point-based segmentation methods to LiDAR point clouds usually leads to unsatisfactory results due to the domain gap between indoor and outdoor scenarios. To address such a domain gap, we propose a new paradigm, namely projected-point-based methods, to transform point-based methods to a suitable form for LiDAR point cloud segmentation by utilizing the characteristics of LiDAR point clouds. Specifically, we utilize the inherent ordered information of LiDAR points for point sampling and grouping, thus reducing unnecessary computation. All computations are carried out on the projected image, and there are only pointwise convolutions and matrix multiplication in projected-point-based methods. We compare projected-point-based methods with point-based methods on the challenging SemanticKITTI dataset, and experimental results demonstrate that projected-point-based methods achieve better accuracy than all baselines more efficiently. Even with a simple baseline architecture, projected-point-based methods perform favorably against previous state-of-the-art methods. The code will be released upon paper acceptance.

Motivation & Objective

• Address the domain gap between indoor and outdoor LiDAR point cloud segmentation, where existing point-based methods underperform in outdoor scenarios.
• Leverage the inherent ordered structure of LiDAR scans to design a more efficient and accurate segmentation framework.
• Develop a new paradigm—projected-point-based methods—that transforms point-based networks for better suitability to LiDAR data.
• Achieve state-of-the-art performance on the SemanticKITTI dataset with minimal architectural complexity.

Proposed method

• Transform 3D LiDAR point clouds into 2D projected images using the inherent angular ordering of LiDAR scans.
• Perform point sampling and grouping based on the projected image structure to reduce computational overhead.
• Conduct all operations—pointwise convolutions and matrix multiplications—on the projected image space rather than raw 3D coordinates.
• Use the projected image representation to maintain spatial relationships while enabling efficient computation.
• Design a lightweight network that operates solely on the projected image, avoiding complex 3D operations.
• Preserve the simplicity of point-based networks while adapting them to exploit LiDAR-specific data ordering.

Experimental results

Research questions

• RQ1 Can a projected-point-based approach effectively bridge the domain gap between indoor and outdoor LiDAR point cloud segmentation?
• RQ2 How does leveraging the inherent ordering of LiDAR scans improve segmentation efficiency and accuracy compared to standard point-based methods?
• RQ3 To what extent can a simple network architecture achieve competitive performance when using projected-point-based computation on outdoor LiDAR data?
• RQ4 Does the projected image representation maintain sufficient geometric information for accurate semantic segmentation in complex outdoor scenes?

Key findings

• Projected-point-based methods achieve higher accuracy than all point-based baselines on the SemanticKITTI dataset.
• The method demonstrates superior efficiency due to reduced computation, relying only on pointwise convolutions and matrix multiplication.
• Even with a simple network architecture, the proposed method outperforms previous state-of-the-art methods on the SemanticKITTI benchmark.
• The use of projected images enables effective exploitation of LiDAR’s inherent scan ordering for improved sampling and grouping.
• The approach effectively mitigates the domain gap between indoor and outdoor segmentation by adapting to outdoor data characteristics.
• The method maintains high performance while minimizing computational complexity, making it suitable for real-time applications.