Single-view 3D Generation via Related Structure Pattern Matching

Some 3D generation techniques have become more general with the advancement of GANs [10] and diffusion models [11]. However, for objects represented by only a single image, depth maps must be estimated using additional information, which makes the collection of training data challenging. In this paper, we propose a framework for establishing 3D generation from a single image by employing a related structure pattern matching approach. First, we extract guided-color and guideddepth components to segment the input image as guidance. These guided components play a critical role in matching related structures, particularly those corresponding to similar pattern images. Second, our method prioritizes targeted pixels by iteratively analyzing matched pattern patches, starting from the farthest patch and moving toward the center until all patches are processed [12]. This method enables real-time transformation at iterative rates for the patterned surfaces of an image, without relying on the strict matching assumptions required for multiple images. This matching approach efficiently identifies and relates pixels and points across several patches.