Continuous Landmark Detection with 3D Queries

Figure 1 . This paper presents a continuous landmark detector that is controlled intuitively with 3D queries. Once trained, a single model can be evaluated with the standard 68 landmarks layout (a), any arbitrarily dense layout (b), and can even be evaluated off-surface to get bone landmarks (c) and (d), which can be used to fit plausible skull and jaw meshes (e) to images. The same network can also be used to infer extremely dense landmarks, enabling applications such as face segmentation (f) or monocular performance capture (g).