Guardians of Hair: Restoring 3D Hair Boundaries
HairGuard is a 3D vision framework for recovering soft boundaries where foreground and background colors mix, such as fine hair. It targets detailed boundary preservation in monocular depth estimation
At fine, translucent boundaries such as hair, foreground and background color information becomes mixed within a single pixel. HairGuard, presented in this paper, is a framework for reconstructing these “soft boundaries” more precisely during depth estimation and view transformation. It addresses issues in existing methods where boundary details disappear, textures in stereo imagery become blurred, and the geometry of new views becomes inconsistent.
The researchers propose a data-construction procedure that uses image matting datasets for training, along with a depth fixer network that automatically identifies soft-boundary regions. It is designed to finely correct depth around these areas through gated residual modules while preserving the overall quality of the scene’s depth, and can be combined with existing state-of-the-art depth models like a plug-in.
During novel-view synthesis, textures are preserved through depth-based forward warping, after which regions that become visible due to disocclusion are filled using a generative scene painter. Finally, the color fuser combines the warped result with the completed image, reducing unwanted background traces around boundaries and producing results with more consistent geometry and finer details.
According to the official description, HairGuard improved performance in soft-boundary regions across several experiments involving monocular depth estimation, stereo image and video transformation, and novel-view synthesis. The paper was released at CVPR 2026 and involved Xiang Zhang, Yang Zhang, Lukas Mehl, Markus Gross, and Christopher Schroers.