Computer Vision News - April 2022

9 DAD-3DHeads Experiment 1) All these metrics are used to compare the state-of-the-art 3D Dense Head Alignment models on the DAD-3DHeads Benchmark. On the full test dataset, the NME for DAD-3DNet for this task is 2.302, significantly lower than 3.580 for the second-best of these methods (3DDFA-V2). Similar results are foundonsubsetsof atypical poses, compound expressions and heavy occlusion. DAD- 3DNet shows superior performance in all cases, which can be also observed in the qualitative results below. From left to right, we find 3DDFA-V2, FaceSynthetic, JVCR, DAD-3DNet. Experiment 3) Similar analysis using all the metrics has been conducted on multiple subgroups (camera pose, age, image quality, occlusions, expressions, lighting) to show robustness of the proposed approach across various conditions (distribution shifts) in-the-wild. Experiment 2) The tasks of 3D Head Face Shape Reconstruction and Head Pose estimation are also evaluated on standard benchmarks (the NoW Face Challenge, and the Feng et al. benchmark) showing advantageous or comparable performance to other state- of-the-art methods. Conclusion Through this work, the authors stress the importance of a diverse, accurate, dense, and in-the-wild dataset for 3D face analysis. The quality of the dataset is boosted by the efficiency of the loss components and the use of ad-hoc evaluation metrics, leading to the conclusion that incorporating information about the full head can improve the model stability, and that the landmarks regression and coarse heatmap estimation modules significantly improve the model performance. We end on a good note for people who want to extend this work and further improve the performance of 3D Landmark localization models… No method is perfect! And even DAD-3DNet fails under challenging situations. Time to work on future advancements and thank the authors for their beautiful illustrations and the decision to make their dataset and models publicly available. See you all next month 