Core Concepts
Proposing a density-guided translator (DGT) improves unsupervised domain adaptive segmentation of 3D point clouds.
Abstract
The content discusses the importance of 3D point cloud segmentation and the challenges faced in adapting synthetic data to real-world scenarios. It introduces a density-guided translator (DGT) to bridge the domain gap and improve segmentation accuracy. The DGT is integrated into a two-stage self-training pipeline named DGT-ST, which outperforms existing methods in synthetic-to-real segmentation tasks. The method is validated on two datasets, SynLiDAR →semanticKITTI and SynLiDAR →semanticPOSS, showcasing significant improvements in mean Intersection over Union (mIoU).
Directory:
Introduction
Abstract
Related Work
Methodology
Preliminaries
Density-guided translator
Category-level adversarial with prototype
Source-aware consistency LaserMix
Experiments
Setup
Comparisons with previous methods
Ablation Studies
Visual Results
Validation of Components
Effect of Hyperparameters
Stats
Experiments on SynLiDAR →semanticKITTI and SynLiDAR →semanticPOSS
DGT-ST achieves 22.7% mIoU gain in SynLiDAR →semanticKITTI
DGT-ST achieves 12.5% mIoU gain in SynLiDAR →semanticPOSS
Quotes
"We propose a density-guided translator (DGT) that directly bridges the domain gap at the input level."
"DGT-ST outperforms state-of-the-art methods, achieving significant mIoU improvements."