Learning Continuous Implicit Representation for Near-Periodic Patterns

AbstractNear-Periodic Patterns (NPP) are ubiquitous in man-made scenes and composed of tiled motifs with appearance differences caused by lighting, defects, or design elements. A good NPP representation is useful for many applications including image completion, segmentation, geometric remapping. But representing challenging because it needs to maintain global consistency (tiled layout) while preserving local variations (appearance differences). Methods trained on general using a large dataset single-image optimization struggle satisfy these constraints, methods that explicitly model periodicity not robust detection errors. To address challenges, we learn neural implicit coordinate-based MLP single optimization. We an input feature warping module periodicity-guided patch loss handle both variations. further improve the robustness, introduce proposal search use multiple candidate periodicities our pipeline. demonstrate effectiveness method more than 500 images building facades, friezes, wallpapers, ground, Mondrian patterns multi-planar scenes.KeywordsNear-periodic patternsNeural representationSingle