LKPM: Large Kernel Point Mamba for 3D Point Clouds

Song Zhao, Shuhua Wang, Xiaobing Zhou

Abstract: The latest 3D point cloud processing research shows that large kernels are critical for improving performance. However, 3D CNNs are limited by their small receptive fields, while directly applying large kernels in Transformers faces challenges of high computational costs. To address this vital challenge, we propose a large kernel module for point cloud serialization and partitioning, called Large Kernel Point Mamba (LKPM). Combining octrees and space-filling curves allows us to serialize unordered point clouds and avoid the high memory overhead associated with the K-Nearest Neighbors (KNN). Based on this, we design a bidirectional and hierarchical State Space Model(SSM) module, which consists of Point-level SSM (PoM) and Patch-level SSM (PaM), to extract local and global features from point clouds efficiently. This design significantly expands the Mamba’s receptive field. Additionally, we introduce a patch shuffling mechanism to mitigate overfitting effectively. Experimental results show that large Mamba kernels are both feasible and efficient in achieving large receptive fields. Under the conditions of not using additional data and training from scratch, our proposed LKPM method achieves state-of-the-art (SOTA) performance among a series of previous SOTA methods, even surpassing some pretraining-based approaches1.

External IDs:dblp:conf/icmcs/ZhaoWZ25