基于KPConv的全局上下文感知点云语义分割方法

Global Context-Aware Point Cloud Semantic Segmentation Method Based on KPConv

摘要: 针对KPConv网络在解码阶段对全局语义信息建模的局限性，提出了一种全局上下文感知点云语义分割模型(GCA-Net)。该模型在解码阶段逐层引入一种轻量化的全局–局部特征引导机制，使局部特征在上采样恢复过程中受到全局语义信息的约束，从而增强输出特征的语义一致性。具体而言，通过全局注意力池化获得全局语义特征表示，采用瓶颈结构增强特征通道间的依赖关系，融合输入点云的局部特征，据此增强全局与局部语义特征的协同表达能力。为了验证所提模型的有效性，在ISPRS数据集上进行了测试，相比于基准网络KPConv，所提GCA-Net的OA和F1分别提高了0.9%和2.7%。

Abstract: To address the limitation of KPConv in modeling global semantic information during the decoding stage, a global context-aware point cloud semantic segmentation model named GCA-Net is proposed. The model incorporates a lightweight global-local feature guidance mechanism into the decoder, enabling global semantic information to effectively constrain local feature recovery during upsampling and enhancing the semantic consistency of the output features. Global semantic representations are constructed through global attention pooling and refined by a bottleneck structure that strengthens inter-channel dependencies, and are subsequently fused with encoder features to improve the collaborative representation of global and local semantics. Experimental results on the ISPRS dataset show that, compared with the baseline KPConv network, GCA-Net achieves improvements of 0.9% in overall accuracy (OA) and 2.7% in F1 score.