.png){kind=icon}
.png){kind=icon}
Return
Multiple Random Observation Strategy for Enhanced ALS Point Cloud Segmentation
.png){kind=icon}
OA
AI
DOI：10.1109/JSTARS.2025.3607404.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Random sampling (RS) is widely used in data-driven large-scale point cloud processing due to its high computational efficiency. However, it suffers from two key limitations: first, the randomness introduced during forward propagation can lead to unstable feature extraction, potentially compromising model performance; second, RS does not consider the spatial structure of point clouds, which may result in the loss of critical information. These issues are particularly prominent in airborne laser scanning (ALS) point clouds, which typically exhibit severe class imbalance and substantial variations in object scales. To address these challenges, we propose a multiple random observation (MRO) framework that leverages the efficiency of RS while capturing spatially complementary features. Building upon this, we introduce the MRO-based feature aggregation module, which integrates features from multiple observations to improve feature extraction stability and enhance segmentation accuracy. Furthermore, we propose the MRO-based downsampling strategy, which identifies informative points by evaluating interobservation feature differences during downsampling, thereby boosting overall model performance. The proposed methods are integrated into several RS-based backbones and evaluated on two representative ALS datasets (i.e., ISPRS and LASDU), demonstrating strong competitiveness compared with current leading approaches.
Keywords:
Airborne laser scanning (ALS)
point cloud
random sampling (RS)
semantic segmentation
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
5.3
Papers:
1.3K
Citations:
3.0W
Organization
View more organizations