Point-by-Point Semantic Segmentation of 3D Indoor Scenes with Few-Shot Based on ADMPTI
Wen-Jing Gao, Zhenxiong Xu, Ya Wang · 2024
This paper addresses the challenging task of semantic segmentation of 3D scenes, particularly in indoor environments. Many existing 3D point cloud segmentation methods face significant issues, including a reliance on large datasets, weak generalization ability, and insufficient learning of discriminative features when performing point-by-point semantic segmentation of 3D point clouds. To address these challenges, we propose ADMPTI, a novel approach that employs a few-shot learning method. Additionally, to further enhance the model's generalization capability, we introduce the DenseBaseLearner. This component enables the model to more accurately measure the similarity between samples, thereby better guiding the segmentation process. Furthermore, to more effectively capture the internal relationships within feature categories, our model incorporates various attention mechanisms. By integrating these attention mechanisms, the model can focus on critical information, thereby improving its ability to extract discriminative features. The effectiveness of our proposed method is demonstrated through significant improvements on the benchmark dataset S3DIS.