Computer Vision News - April 2022

5 DAD-3DHeads To obtain this dataset, a novel annotation method is used, which employs a 3D modelling tool. A 3D Morphable model is fit to a human head image by picking anchoring points on the 3D mesh surface (head on the left). Based on the pinned points, the tool automatically modifies the mesh to optimise the 3DMM parameters, so that the “pin” reprojection error is minimized. Finally, the annotator can check the texture rendered onto the 3D mesh to verify that it is realistic. The total dataset is made of 42,130 images from different sources. Each of these is provided with 5,023 vertices of which 3,669 (head subset) are proven to be accurately labelled. The flexibility of this dataset is given primarily by the extension of the landmarks, and one can choose from different subsets (see figure on the left: starting with 68 landmarks -> 191 -> 445). The annotation accuracy can be observed qualitatively in the image below, where a) shows the ground truth scan provided by previous datasets, b) is the annotated result, while d) shows the alignment of the mesh and the GT scan.