EZ-SP: Fast and Lightweight Superpoint-Based 3D Segmentation

2025-11-29

AI Summary

Existing superpoint-based 3D semantic segmentation methods suffer from CPU-intensive partitioning steps, hindering real-time inference. This paper proposes the first fully GPU-differentiable superpoint generation framework: it eliminates hand-crafted features and CPU-based preprocessing, and introduces a lightweight, end-to-end trainable superpoint partitioning module. A differentiable proxy loss is designed to optimize superpoint semantic consistency, and a lightweight classifier is tightly coupled for joint optimization. The method achieves state-of-the-art point-wise accuracy on S3DIS, KITTI-360, and DALES, while accelerating inference by 72×, reducing model parameters by 120×, and consuming less than 2 MB of GPU memory. These advances significantly advance the feasibility of real-time, large-scale point cloud semantic segmentation.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Semi-Supervised Learning
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Superpoint-based pipelines provide an efficient alternative to point- or voxel-based 3D semantic segmentation, but are often bottlenecked by their CPU-bound partition step. We propose a learnable, fully GPU partitioning algorithm that generates geometrically and semantically coherent superpoints 13$ imes$ faster than prior methods. Our module is compact (under 60k parameters), trains in under 20 minutes with a differentiable surrogate loss, and requires no handcrafted features. Combine with a lightweight superpoint classifier, the full pipeline fits in $<$2 MB of VRAM, scales to multi-million-point scenes, and supports real-time inference. With 72$ imes$ faster inference and 120$ imes$ fewer parameters, EZ-SP matches the accuracy of point-based SOTA models across three domains: indoor scans (S3DIS), autonomous driving (KITTI-360), and aerial LiDAR (DALES). Code and pretrained models are accessible at github.com/drprojects/superpoint_transformer.

Problem

• Accelerates superpoint generation for 3D segmentation via GPU partitioning
• Reduces model size and memory usage for large-scale 3D scene processing
• Maintains accuracy across diverse domains while enabling real-time inference

Innovation

• GPU-based superpoint partitioning algorithm
• Compact module with under 60k parameters
• Lightweight pipeline for real-time inference