Computer Vision News - April 2022

4 CVPR-Accepted Research Paper DAD-3DHEADS: A LARGE-SCALE DENSE, ACCURATE AND DIVERSE DATASET FOR 3D HEAD ALIGNMENT FROM A SINGLE IMAGE by Marica Muffoletto (twitter) Hi everyone and welcome to a special review this month! We have a paper called DAD-3DHeads: A Large-scale Dense, Accurate and Diverse Dataset for 3D Head Alignment from a Single Image, recently accepted to CVPR 2022. Ralph and I are especially happy with choosing this paper, because it enables us to show support to the main author’s homeland and people . We deeply thank all authors ( Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Viktoriia Sharmanska, Jiří Matas ) for allowing us to use their images. Introduction of a new dataset As suggested by the title, themain contribution of this paper is the DAD-3DHeads Dataset for 3D face analysis. Needless to say, many tasks in Computer Vision rely highly on the quality and quantity of the data employed, and therefore such a contribution might be essential to advance the state-of-the-art. This dataset aims at containing a variety of extreme poses, facial expressions, challenging illuminations, and severe occlusion cases. Moreover, as a big step forward, it proposes to include annotations of 3D landmarks directly from images, that are tested for accuracy and consistency compared to 4D ground truth scans. The authors argue that the lack of those and hence the employment of 2D landmarks in current datasets is one of the main challenges in the training of pose estimation and head alignment models.