Reproduce “img2pose: Face Alignment and Detection via 6DoF, Face Pose Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vítor Albiero, Xingyu Chen, Xi Yin, Guan Pang, Tal Hassner
Researchers from the University of Notre Dame and Facebook AI present img2pose, a method for direct 6DoF face pose estimation that eliminates the need for prior face detection or landmark localization. It achieves state-of-the-art pose accuracy on AFLW2000-3D and BIWI datasets while deriving competitive face detection performance as a byproduct, running in real-time.
Key insights
- Estimating 6DoF rigid transformation is a more direct and informative problem for face alignment than traditional facial landmark detection, offering a 6D regression target compared to 10D or 136D for landmarks.
- Accurate face detection bounding boxes can be effectively derived as a byproduct of estimated 3D face poses, eliminating the need for a separate bounding box prediction head during training.
- A two-stage detection framework like Faster R-CNN can be effectively adapted for direct 6DoF pose estimation by training its Region Proposal Network (RPN) with bounding boxes projected from 6DoF ground-truth pose labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.