The overview of our proposed method. The left side shows the attributes autoencoder, and the right side shows the entropy model. “SConv n³ × C” and “TSConv n³ × C “ denotes the sparse convolution and transposed convolution with C output channels and kernel size n³. “Residual Block” and “Self Attention Block” represent the residual network and the local self attention network used for efficient latent feature aggregation. “s ↑” and “s ↓” represent upsampling and downsampling at a factor of s. “Q” represents quantizer, “AE” represent arithmetic encoder, and “AD” represent arithmetic decoder. “G” represents the partition operation of the quantized latent representations. “0” symbolizes the context with the same shape as the input point cloud, where all attribute values are 0. U is used to describe the combination of point clouds of different shapes. Red arrows represent the encoding data flow, blue arrows represent the decoding data flow, and purple arrows represent the shared data flow.