Computer Vision News - April 2022

7 DAD-3DHeads Here, is a Landmark Regression Loss while is a Gaussian Heatmap Loss. The other two terms are novel: is a Shape+Expression Loss , which disentangles shape and expression frompose,measuring thediscrepancy between thenormalised subsampled (only head) predicted vertices and the ground truths in 3D. ( Reprojection Loss ) optimises the pose accuracy through a projection of the 3D vertices of the mesh into the image. The L1 loss is used to measure the difference between the reprojected subsampled vertices. The encoder is pre-trained on ImageNet. The FLAME layer is fixed during training, and ADAM optimiser with learning rate = 1x10 -4 which reduces when validation loss stops decreasing. No image augmentation is used. Experiments focus on: 1. Evaluate the task of 3D Dense Head Alignment from an image 2. Test model generalization (on a range of tasks - Face Shape Reconstruction, Head Pose Estimation) 3. Test robustness to extreme poses Evaluation Protocol In this paper, two new metrics for 3D Head Learning tasks are introduced. These are the Reprojection NME and the accuracy . The former computes the normalised mean error (NME) of the reprojected 3D vertices onto the image plane for 68 landmarks. The latter finds K/N closest vertices in the ground-truth mesh and calculates which of them is closer or further from the camera, finally it compares to the configuration for the predicted vertices . Moreover, traditional metrics are employed: Chamfer Distance , for measuring the accuracy of fit on any number of predicted vertices and Pose error to measure accuracy of pose prediction based on rotation matrices.