Computer Vision News - April 2022

6 CVPR-Accepted Research Paper The quality of the 3D novel annotations is also quantitatively measured through comparison with the manually labelled facial landmarks (found by 10 different annotators). This is done by computing the quality score for each approach averaging across the images as a normalised mean error between each pair of labels, resulting in an average NME 45.8% lower in the 3D method compared to the manual one. where N is a subset of 30 images, is the head bounding box size, and is an array of 68 labelled landmarks. Experiments with the new dataset To demonstrate the efficiency of the dataset, a data-driven DAD-3DNet model is trained. This is made of a CNN encoder , a Landmark Heatmap Estimator that predicts coarse locations of 2D landmarks, a Fusion Module that fuses the heatmap prediction with the CNN encoder features, and a RegressionModule that predicts finer facial landmark’s locations and 3DMM parameters vectors. FLAME Layer maps the 3DMM vector to 3D head model vertices. The overall loss employed is a combination of 4 terms with respective weights: