Deformed Implicit Field: Modeling 3D Shapes With Learned Dense Correspondence

Figure 1. Our DIF-Net can produce 3D shapes with dense correspondences for object categories containing complex geometry variation and structure differences. It enables high-quality texture transfer shown in the middle four columns, where the two smaller figures after each transfer result show the color-coded correspondences (top) and their uncertainty (bottom; blue and red indicates low and high uncertainty respectively). With our learned shape space and correspondence, shapes can be freely edited by simply moving one or a sparse set of points, as shown in the last two columns.