Semantic-aware DropSplat: Adaptive Pruning of Redundant Gaussians for 3D Aerial-View Segmentation

2025-08-13

AI Summary

To address semantic ambiguity in 3D aerial scenes caused by scale variation and structural occlusion, this paper proposes a semantics-aware adaptive pruning method for Gaussian points. The method employs Gaussian point clouds as the geometric representation and introduces two key innovations: (1) a learnable sparsification mechanism based on the Hard Concrete distribution, jointly optimized with a semantic confidence estimation module to dynamically remove redundant Gaussian points; and (2) a weakly supervised optimization paradigm leveraging high-confidence pseudo-labels generated by a pre-trained 2D foundation model. By preserving semantic consistency while significantly enhancing representational compactness, the approach achieves superior segmentation accuracy and inference efficiency under limited annotations. Experiments on our newly constructed 3D-AS dataset demonstrate substantial improvements in both precision and computational efficiency, enabling high-fidelity, compact semantic understanding of complex 3D aerial scenes.

Technology Category

• Computer Vision: 3D Computer Vision
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Semantics and Knowledge: Methods, algorithms and applications for the development of semantic models, knowledge graphs and other forms of structured data models with machine-interpretable semantics
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

In the task of 3D Aerial-view Scene Semantic Segmentation (3D-AVS-SS), traditional methods struggle to address semantic ambiguity caused by scale variations and structural occlusions in aerial images. This limits their segmentation accuracy and consistency. To tackle these challenges, we propose a novel 3D-AVS-SS approach named SAD-Splat. Our method introduces a Gaussian point drop module, which integrates semantic confidence estimation with a learnable sparsity mechanism based on the Hard Concrete distribution. This module effectively eliminates redundant and semantically ambiguous Gaussian points, enhancing both segmentation performance and representation compactness. Furthermore, SAD-Splat incorporates a high-confidence pseudo-label generation pipeline. It leverages 2D foundation models to enhance supervision when ground-truth labels are limited, thereby further improving segmentation accuracy. To advance research in this domain, we introduce a challenging benchmark dataset: 3D Aerial Semantic (3D-AS), which encompasses diverse real-world aerial scenes with sparse annotations. Experimental results demonstrate that SAD-Splat achieves an excellent balance between segmentation accuracy and representation compactness. It offers an efficient and scalable solution for 3D aerial scene understanding.

Problem

• Addresses semantic ambiguity in 3D aerial segmentation
• Eliminates redundant Gaussian points for compact representation
• Improves segmentation accuracy with limited ground-truth labels

Innovation

• Semantic-aware Gaussian pruning with Hard Concrete
• High-confidence pseudo-labels from 2D foundation models
• Novel 3D aerial benchmark dataset with sparse annotations