OctFormer: Efficient Octree-Based Transformer for Point Cloud Compression with Local Enhancement

Point cloud compression with a higher ratio and tiny loss is essential for efficient data transportation. However, previous methods that depend on 3D convolution or frequent multi-head self-attention operations bring huge computations. To address this problem, we propose an octree-based Transformer method called OctFormer, which does not rely the occupancy information of sibling nodes. Our uses non-overlapped context windows to construct octree node sequences share result operation among sequence Besides, introduce locally-enhance module exploiting features positional encoding generator enhancing translation invariance sequence. Compared state-of-the-art works, our obtains up 17% Bpp savings compared voxel-context-based baseline saves overall 99% coding time attention-based baseline.