Training-Free Point Cloud Recognition Based on Geometric and Semantic Information Fusion
Yan Chen, Di Huang, Zhichao Liao, Xi Cheng, Xinghui Li, Long Zeng
TL;DR
This work tackles the need for efficient point cloud recognition without training by integrating geometric and semantic cues in a training-free dual-branch framework. It combines non-parametric geometric processing with a text-aligned semantic encoder (ULIP), powered by a feature memory (MEM) and a weighted fusion $f_{fuse} = \alpha f_{geo} + (1 - \alpha) f_{sem}$. Key contributions include Memory Feature Filtering via K-Means++ to form a compact $F_{key}$, Geometric Feature Enhancement that adds spherical coordinates and edge-based features, and extensive experiments showing strong performance on ModelNet and ScanObjectNN in both full-shot and few-shot settings. The approach offers practical benefits for resource-constrained deployment by delivering accurate recognition without parameter updates during inference.
Abstract
The trend of employing training-free methods for point cloud recognition is becoming increasingly popular due to its significant reduction in computational resources and time costs. However, existing approaches are limited as they typically extract either geometric or semantic features. To address this limitation, we are the first to propose a novel training-free method that integrates both geometric and semantic features. For the geometric branch, we adopt a non-parametric strategy to extract geometric features. In the semantic branch, we leverage a model aligned with text features to obtain semantic features. Additionally, we introduce the GFE module to complement the geometric information of point clouds and the MFF module to improve performance in few-shot settings. Experimental results demonstrate that our method outperforms existing state-of-the-art training-free approaches on mainstream benchmark datasets, including ModelNet and ScanObiectNN.