Reproduce “CLERF: Contrastive LEaRning for Full Range Head Pose Estimation” on OpenResearch

About this paper

Ting-Ruen Wei, Haowei Liu, Huei-Chung Hu, Xuyang Wu, Yi Fang, Hsin-Tai Wu

CLERF introduces a contrastive learning framework for full-range head pose estimation, leveraging 3D-aware GANs to generate synthetic positive samples and geometric transformations to expand data coverage. The model achieves robust 360-degree head pose predictions for all Euler angles, demonstrating a Mean MAE of 4.98 degrees on fully augmented AFLW2000 and outperforming existing methods by over 10 degrees under extreme transformations.

Key insights

• Synthetic generation of anchor-positive samples using 3D-aware GANs effectively addresses data sparsity, making contrastive learning feasible for full-range head pose estimation.
• Geometric transformations, including rotations and flips, can rigorously expand training data for full 360-degree pose coverage and enhance robustness, while a mathematical theorem ensures the integrity of contrastive learning triplets.
• Contrastive learning, applied with an appropriate loss function and image encoder, enables the model to learn more discriminative head pose features, improving accuracy for fine-grained angular differences and robustness to transformations.

Summary by alphaXiv.