Shape Alignment by Learning a Landmark-PDM Coupled Model

This paper revisits the model-based approaches for groupwise shape alignment. The key contribution is modeling the landmarks instead of considering them as nodes sliding along the shape contour. The shape group is thus modeled by a landmark-PDM coupled model instead of a constrained point distribution model (PDM). This coupled model is estimated by a stable four-stage estimation algorithm. There are two significant achievements. First, shapes are aligned in a fully unsupervised manner - both the number and location of landmarks are automatically decided. Second, extremely noisy and largely deformed shapes can be robustly aligned. These are demonstrated using both synthesized and real data