Reproduce “EFHQ: Multi-purpose ExtremePose-Face-HQ dataset” 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
Trung Tuan Dao, Duc Hong Vu, Cuong Pham, Anh Tran
The EFHQ dataset, developed by VinAI Research, introduces a multi-purpose collection of up to 450,000 high-quality facial images with extreme head poses to address the frontal-view bias in existing datasets. This resource improves deep learning model performance in face generation, reenactment, and recognition tasks involving diverse facial orientations.
Key insights
- An ensemble method for head pose estimation significantly increases robustness and accuracy in classifying extreme facial orientations, overcoming limitations of single estimators.
- State-of-the-art face recognition models exhibit notable performance drops (5-37% TAR@FAR=1e-3) when evaluated on cross-pose verification benchmarks, exposing a critical vulnerability.
- Augmenting training data with high-quality, identity-annotated extreme-pose images substantially improves generalization and performance across various facial tasks without compromising frontal-view capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.