FPNforV-PCC

This dataset is for paper "A feature pyramid network based partition map prediction method for efficient encoding in Video-based Point Cloud Compress", and is what the authors use for training the FPN mentioned in the paper. The CU partitioning data in this dataset comes from V-PCC using VVC to make CU partition and extracting data for a sample based on 64×64 CU. Encoder configuration when extracting data as described in Section III-A of the paper, the point cloud source is the first 32 frames of basketball_player order by owlii, including the partitioning of 5 different QPs. In the dataset, "AttrI_TOri" and "GeomP_TOri" represents the input for luma values, "AttrI_Resi" and "GeomP_Resi" stands for the prediction residual input, "AttrI_Occu" and "GeomP_Occu" denotes the placeholder information input, and "AttrI_y_QP" and "GeomP_y_QP" encompasses both the label value and QP input. The program source code is available on https://github.com/Mesks/FPNforV-PCC/tree/master.

本数据集旨在服务于论文《基于特征金字塔网络的视频点云压缩高效编码中的分区图预测方法》，并为文中提及的FPN（特征金字塔网络）的训练提供了依据。数据集中的CU（宏块）分区数据源自V-PCC，通过VVC进行CU分区，并基于64×64 CU提取样本数据。提取数据时遵循论文第三章A节所述的编码器配置，点云源为owlii按顺序排列的篮球运动员的前32帧，包括5种不同质量点（QP）的分区。在数据集中，'AttrI_TOri'与'GeomP_TOri'代表亮度值的输入，'AttrI_Resi'与'GeomP_Resi'表示预测残差输入，'AttrI_Occu'与'GeomP_Occu'表示占位符信息输入，而'AttrI_y_QP'与'GeomP_y_QP'则包含了标签值与QP输入。程序源代码可在https://github.com/Mesks/FPNforV-PCC/tree/master上获取。