Kernel-Based Attention Network for Point Cloud Compression.

LiDAR point clouds are widely adopted in robotics to precisely depict the surrounding environment. However, dense point cloud acquisition significantly increases data processing costs. To address the extensive memory demands of dense 3D maps, compression techniques are crucial for efficient storage and transmission. Traditional point cloud compression methods overlook local structural information and contextual relationships among point clouds. To tackle this issue, we propose a kernel-based attention network, which employs the attention mechanism to capture contextual information. Our novel deep convolutional encoder directly operates on the point cloud, avoiding structural information loss caused by voxelization. Furthermore, we introduce a self-attention feature aggregation module for extracting features and obtaining more robust representations. We evaluate our method on KITTI and nuScenes datasets. Experimental results show that our compression network achieves superior reconstruction compared to other state-of-the-art approaches at the same bit-rate.