||An Encoding Method of Multi-view Images using Plenoptic Point Cloud
||백무현(Mu Hyun Back) ; 문지원(Ji Won Moon) ; 이진규(Jin Kyu Lee) ; 이하현(Hahyun Lee) ; 강정원(Jungwon Kang) ; 황성수(Sung Soo Hwang)
|| Multi-view; plenoptic point cloud; 3D Reconstruction; Voxelization
||In this paper, we propose a system to create a plenoptic point cloud from multi-view and depth images and efficiently encode multi-view images using plenoptic point cloud. Plenoptic Point Cloud is a set of plenoptic points. Plenoptic Point is an representation method that has single geometric information where 3D point is located and multiple color information obtained from multi-view images. With the recent increase in interest in immersive contents, the utilization of multi-view images has increased, and research on encoding is essential to efficiently deliver them. To solve this, this paper uses plenoptic point cloud to minimize the loss of color information, and performs cube and frustum-based voxelization. Then, to solve the problem of geometric information distortion caused by voxelization, we propose a method to regenerate an image by projecting it in consideration of camera position. At the same time, similarity analysis between planoptic points is performed to verify whether this paper proposed an expression method advantageous for encoding. Experiment results showed that the average PSNR of images generated by creating a plenoptic point cloud using incremental generation method for all datasets and voxelizing them into frustum with a resolution of 4096 were 50.7dB. In addition, high similarities between plenoptic points of similar positions leads to an advantage in compression and encoding.