LKG-Net: Local-Keypoint-Global Feature Fusion Network for Sparse Point Cloud Semantic Segmentation

Semantic segmentation of sparse point clouds presents a significant challenge in 3D computer vision. Sparsity primarily arises from two sources: suboptimal data quality due to limitations in point cloud acquisition devices or environmental factors, and weakly supervised learning strategies employed to reduce annotation costs. Existing methods often struggle to simultaneously achieve accurate perception of fine-grained local geometric structures and effective understanding of global scene context when processing sparse point clouds. To address this challenge, we propose LKG-Net, a novel end-to-end segmentation network that systematically enhances feature learning and fusion capabilities through three carefully designed modules. The Adaptive Hierarchical Efficient MLP Module (AHEM) enables deep extraction of robust local features through hierarchical feature refinement and adaptive pooling strategies. The Keypoint-Driven Local-to-Global Spatial Feature Module (KGSF) employs an efficient keypoint-based attention mechanism to capture global contextual information while significantly reducing computational complexity. The Local-Global Fusion Module (LGF) dynamically and adaptively merges multi-scale features based on data characteristics, ensuring optimal feature integration across different spatial scales. Comprehensive experiments on the S3DIS and Toronto-3D benchmark datasets demonstrate the effectiveness of our method. Specifically, LKGNet establishes a new state of the art on the S3DIS Area 5 test set under an extremely sparse (0.1 %) supervision setting, achieving a mean Intersection over Union (mIoU) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 0. 4 \%}$</tex> and surpassing strong baselines by 2.7 %. On Toronto-3D, LKG-Net further demonstrates robust generalization with superior performance on complex outdoor scenes. These results conclusively establish LKG-Net's exceptional capability for precise segmentation under sparse point cloud conditions, providing a robust and generalizable solution for sparse supervised 3D scene understanding.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.