A Density-Aware Point Cloud Geometry Compression Leveraging Cluster- Centric Processing
Fariha Afsana,Manoranjan Paul,Faranak Tohidi,Pan Gao
DOI: https://doi.org/10.1109/access.2024.3411029
IF: 3.9
2024-06-15
IEEE Access
Abstract:Recent years have encountered a noticeable expansion of point cloud-based 3D applications that compels the necessity of high-efficiency point cloud compression. Preserving the local density of point cloud is crucial for compression, however, it's challenging due to the unordered nature of points in 3D space and has been overlooked by the majority of the existing compression methods. Recently, the farthest point method has demonstrated effectiveness in sampling point clouds across various state-of-the-art methods, its limitation lies in being non-density aware, yielding reconstructions that fall short of the desired quality. Intending to achieve density-aware compression for improved reconstruction, this paper proposes an end-to-end learnable cluster-based compression method for efficient lossy point cloud geometry compression. Our method utilizes an autoencoder architecture that performs point-wise operation for compression and reconstruction. To extract effective latent features, the encoder segments the point clouds into clusters using a different approach than farthest point sampling that is capable of well capturing uneven point densities and employs cluster-wise compression independently. To enhance density retention further, we leveraged the capabilities of an attention mechanism, allowing it to learn complex point-wise dependencies within clusters and effectively capture local density information. At the decoder, decompressed clusters are accumulated to reconstruct point clouds completely. In terms of rate-distortion trade-off, the experimental analysis reveals the superiority of the proposed method over prior arts. Furthermore, our approach adapts well to different datasets, for example, the model learned from the ModelNet40 dataset works well and achieves state-of-the-art performances on ShapeNet datasets. Finally, the qualitative comparison demonstrates that the proposed framework can preserve satisfactory local geometry details of point clouds compared to existing Representative methods.
computer science, information systems,telecommunications,engineering, electrical & electronic