[Paper Review] 6D Rotation Representation For Unconstrained Head Pose Estimation

Thorsten Hempel, Ahmed A. Abdelrahman+1 more|arXiv (Cornell University)|Feb 25, 2022

Medical Imaging and Analysis4 citations

TL;DR

This paper proposes 6DRepNet, a landmark-free head pose estimation method that uses a continuous 6D rotation matrix representation to enable full-range, unambiguous 6D rotation regression. By training with a geodesic loss that respects SO(3) manifold geometry, it achieves state-of-the-art performance, outperforming prior methods by up to 20% on AFLW2000 and BIWI datasets.

ABSTRACT

In this paper, we present a method for unconstrained end-to-end head pose estimation. We address the problem of ambiguous rotation labels by introducing the rotation matrix formalism for our ground truth data and propose a continuous 6D rotation matrix representation for efficient and robust direct regression. This way, our method can learn the full rotation appearance which is contrary to previous approaches that restrict the pose prediction to a narrow-angle for satisfactory results. In addition, we propose a geodesic distance-based loss to penalize our network with respect to the SO(3) manifold geometry. Experiments on the public AFLW2000 and BIWI datasets demonstrate that our proposed method significantly outperforms other state-of-the-art methods by up to 20\%. We open-source our training and testing code along with our pre-trained models: https://github.com/thohemp/6DRepNet.

Motivation & Objective

• Address the ambiguity and discontinuity issues in Euler angle and quaternion representations for head pose estimation.
• Overcome the limitations of binning-based classification methods that restrict pose prediction to narrow angular ranges.
• Enable end-to-end direct regression of full 6D head pose using a continuous, unambiguous rotation matrix representation.
• Improve training stability and generalization by leveraging SO(3) manifold geometry through a geodesic distance-based loss function.
• Demonstrate superior performance across diverse datasets without relying on landmark detection or angle binning.

Proposed method

• Represent 3D rotations using a 6D compressed form of the 3×3 rotation matrix, derived by dropping the third column vector to avoid enforcing orthogonality during training.
• Map the predicted 6D vector back into SO(3) via Gram-Schmidt orthogonalization to reconstruct a valid rotation matrix.
• Use geodesic distance as the loss function to penalize prediction errors in the intrinsic Riemannian geometry of SO(3), improving robustness over standard MSE.
• Train the network end-to-end using a deep CNN backbone (e.g., RepVGG or ResNet50) to regress the 6D representation directly from image input.
• Avoid discretization of rotation space by eliminating the need for classification into angular bins, enabling full-range regression.
• Leverage the rotation matrix’s unique parameterization to eliminate ambiguity and improve generalization across extreme head poses.

Experimental results

Research questions

• RQ1 Can a continuous 6D rotation representation improve full-range head pose estimation compared to discrete binning or Euler/quaternion representations?
• RQ2 Does training with a geodesic loss that respects SO(3) manifold geometry lead to more accurate and robust pose regression than standard MSE loss?
• RQ3 Can direct regression of a 6D rotation representation outperform landmark-based or binning-based methods in unconstrained settings?
• RQ4 How does the choice of backbone network affect the performance of the proposed 6D rotation regression framework?
• RQ5 Can the proposed method generalize across diverse datasets without performance degradation on extreme or unconstrained poses?

Key findings

• The proposed 6DRepNet method achieves a mean absolute error (MAE) of 2.66 on the BIWI dataset using geodesic loss, outperforming all prior state-of-the-art methods.
• On the AFLW2000 dataset, the method reduces MAE to 3.97 with geodesic loss, representing a 20% improvement over previous SOTA methods.
• Training with geodesic loss yields slightly better performance than with ℓ₂ loss, demonstrating the advantage of manifold-aware optimization.
• The RepVGG backbone achieves 3.97 MAE on AFLW2000 and 2.66 MAE on BIWI, outperforming ResNet50 by ~7% across all test scenarios.
• The method maintains consistent performance across yaw, pitch, and roll angles, indicating robustness to full-range pose variations.
• The ablation study confirms that the geodesic loss improves generalization and reduces error accumulation in high-dimensional rotation space.