CEM-DIT: Context Entropy Model with Dual Interactive Transformer for Point Cloud Geometry Compression

Recent advancements in Point Cloud Geometry Compression (PCGC) have seen the emergency of Transformer-based architectures. These networks utilize interactions across various dimensions such as spatial or channel domain, achieving impressive performance. Inspired by this, we propose a novel Context Entropy Model with Dual Interactive Transformer (CEM-DIT) that combines both dimensions to enhance representation capability in PCGC. Specifically, our CEM-DIT model predicts the occupancy code distribution of each octree node by aggregating features across spatial and channel dimensions through intra-block interactions. The architecture incorporates a Dual Interactive Spatial Attention (DISA) module in consecutive Transformer blocks, comprising a Sliding Context Window Self-Attention (SCW-SA) block for capturing similarity and strong dependency among the nodes in the spatial dimension, and a Dual Feature Interaction (DF-I) block for inferring spatial and channel attention maps. Furthermore, we introduce a Convolutional Aggregation based Spatial-gate Feed-forward (CASF) module to achieve intra-block feature aggregation from the ancestors and siblings, incorporating non-linear spatial information into the feed-forward network. Extensive experimental results demonstrate that our proposed CEM-DIT outperforms current state-of-the-art PCGC methods.