Artem Sevastopolsky and Dmitrii Pozdeev - DenseMarks Learning Canonical Embeddings for Human Heads

This session will present DenseMarks, a novel learned representation for human heads that enables high-quality dense correspondences across head images. The work introduces a Vision Transformer-based model that predicts a 3D embedding for every pixel of a 2D head image, mapping it into a canonical 3D unit cube where semantically corresponding regions align across identities, poses, and appearances. The authors construct a large-scale supervision signal by extracting pairwise point matches from diverse in-the-wild talking head videos using a state-of-the-art point tracker, and train the model with a contrastive objective that brings embeddings of matched points closer together in canonical space. To further structure this space, the method incorporates multi-task learning with facial landmarks, head segmentation constraints, and latent cube features that enforce spatial continuity, resulting in an interpretable and queryable canonical embedding space. For the Cohere Labs computer vision community, this session will highlight how DenseMarks supports downstream tasks such as discovering common semantic parts, robust face and head tracking, and stereo reconstruction, while remaining resilient to large pose changes and capturing the full head including hair. Attendees will learn how the canonical space bottleneck leads to pose- and identity-consistent representations and will see state-of-the-art results in geometry-aware point matching and monocular head tracking with 3D Morphable Models. Artem is currently a research engineer at Apple, Munich Vision Lab. Previously, he was a Ph.D. student at TUM (Visual Computing & AI lab led by prof. Matthias Nießner). Prior to that, he worked as a deep learning engineer at the Vision, Learning and Telepresence lab at Samsung AI Center Moscow and studied at Skoltech under supervision of prof. Victor Lempitsky. He has recently been a research intern at Meta Reality Labs, Codec Avatars Team. At Samsung & Skoltech he worked on realistic