Neural Surface Maps

Figure 1 : Here, we show the extracted map between two non-isometric shapes, tiger and iguana. Although the shapes are highly nonisometric, e.g., lengths of the tail and legs, our method successfully associated these regions between shapes, thus yielding a semantically correct map. This map is seamless by construction, and optimized with no supervision thanks to pre-trained ViT models which can identify semantically corresponding points across shape renderings.