(A) study on voxel texture compression based on surface traversal scanning

Abstract

Recently, the three-dimensional point cloud is easily acquired by depth sensors, LiDAR, stereo cameras and they have been utilized in computer vision research fields, e.g., autonomous driving, immersive media and AR, and so on. To utilize point clouds in various applications, they are usually converted to voxel data. As voxel data is used in many applications, the encoding of voxel data became a crucial task. Voxel data consists of geometry and texture data, and texture data is an important component for visualizing voxel data. Hence, many studies have focused on encoding voxel texture. In particular, texture atlases are widely used for representing voxel texture. The texture atlas is typically utilized for encoding voxel texture using image/video encoders, so it is important to design a texture atlas while maintaining the connectivity between voxels. To achieve this goal, MPEG introduces video-based point cloud compression which is the state of the art. In video-based point cloud compression, voxel data is segmented into a set of patches, and a texture atlas is generated to project the patches onto two-dimensional planes. However, the texture atlas generated from video-based point cloud compression is likely to lose the connectivity of voxels because of segmentation. In this dissertation, I present a novel voxel texture map generation method for encoding voxel texture by using surface traversal scanning scheme. To scan voxel texture, a boundary tracing algorithm is utilized. Boundary tracing algorithms identify the successive boundary pixels on binary images, accordingly making a useful tool for scanning voxel texture while keeping coherence of voxels. To utilize a boundary tracing algorithm to voxel data, voxel data is decomposed as a set of binary images that I named as slices, and voxel texture is scanned for each slice. Also, I present a voxel texture map that is generated using scanned texture arrays. Because the scanned texture arrays are generated from slices, adjacent texture arrays are highly correlated. Hence, a texture atlas is generate using aligning scanned voxel texture. Besides, I also proposed a partitioning method for a voxel data to solve the self-occlusion problem that occurs in the alignment process. To encode the proposed texture atlas, the image/video encoders are utilized and I proposed encoding strategies. Experimental results showed that the proposed method outperforms the previous methods.