Supplementary Materials

Tables 1, 2 and 3 fully demonstrate the architecture of the proposed ADNet. For detailed introduction of our experimental setting, please refer to Section 4 of our manuscript. In the table, P∗, H∗point and H∗edge denote the inputs of smooth ADL1 loss, AWing loss and AWing loss, respectively. Npoint and Nedge indicate the number of points and edges, which varies according to each dataset. The loss weights of Hour Glass (HG) for stacked 4 HGs are respectively 1/8, 1/4, 1/2, and 1. The fourth head branch outputs P3 is the final predicted coordinate of each landmark, which is derived from the soft argmax operation. In Table 2, the goal of E2P Transform is to convert Ĥedge (Nedge channels) into Hedge (Npoint channels) by considering the adjacency relationship as