HybridPose: 6D Object Pose Estimation under Hybrid Representations (2001.01869v4)

Abstract: We introduce HybridPose, a novel 6D object pose estimation approach. HybridPose utilizes a hybrid intermediate representation to express different geometric information in the input image, including keypoints, edge vectors, and symmetry correspondences. Compared to a unitary representation, our hybrid representation allows pose regression to exploit more and diverse features when one type of predicted representation is inaccurate (e.g., because of occlusion). Different intermediate representations used by HybridPose can all be predicted by the same simple neural network, and outliers in predicted intermediate representations are filtered by a robust regression module. Compared to state-of-the-art pose estimation approaches, HybridPose is comparable in running time and accuracy. For example, on Occlusion Linemod dataset, our method achieves a prediction speed of 30 fps with a mean ADD(-S) accuracy of 47.5%, representing a state-of-the-art performance. The implementation of HybridPose is available at https://github.com/chensong1995/HybridPose.

• The paper introduces a hybrid approach combining keypoints, edge vectors, and symmetry correspondences to robustly predict 6D object poses.
• It employs dual modules—Prediction and Pose Regression—with specialized neural networks to address challenges like occlusion and outlier management.
• Empirical results report 91.3% ADD(-S) on Linemod and 47.5% on Occlusion Linemod, with real-time performance at 30 fps on standard GPUs.

HybridPose: 6D Object Pose Estimation under Hybrid Representations

In the paper "HybridPose: 6D Object Pose Estimation under Hybrid Representations," the authors present a novel approach for estimating the 6D pose of objects from RGB images. This study addresses the fundamental challenge in 3D vision of predicting object position and orientation using hybrid intermediate representations. HybridPose stands out by incorporating keypoints, edge vectors, and symmetry correspondences, offering enhanced robustness and accuracy compared to methods relying on unitary representations.

Methodology Overview

HybridPose introduces a framework with two main modules: Prediction and Pose Regression.

1. Prediction Module: It employs three neural networks to independently predict:
   • Keypoints: Standard in previous works, they provide sparse geometric information.
   • Edge Vectors: These capture spatial relationships, refining constraints between adjacent keypoints.
   • Symmetry Correspondences: Leveraging object symmetries, they enhance accuracy, particularly under occlusion by constraining rotational parameters in the pose.
2. Pose Regression Module: This module refines the pose estimates using robust norms to handle outliers in the predicted elements, an improvement over traditional methods which may struggle with inaccuracies introduced by occlusions.

Key Results

The paper evaluates HybridPose on the Linemod and Occlusion Linemod datasets. Results indicate:

• On Linemod, HybridPose achieves an ADD(-S) accuracy of 91.3%, demonstrating its efficacy in accurately estimating poses.
• On Occlusion Linemod, it surpasses existing methods with an accuracy of 47.5%, showing superior performance in occluded scenarios.

The efficiency is also notable, with the system running at 30 fps on standard GPU hardware, making it suitable for real-time applications.

Implications and Future Directions

HybridPose demonstrates that integrating multiple intermediate representations can significantly improve pose estimation accuracy, particularly in challenging conditions like occlusions. This hybrid approach offers a promising direction for advancing object recognition and manipulation tasks in robotics.

Future research could explore incorporating additional representations like shape primitives or normals and enforcing cross-representation consistency, potentially enhancing the robustness and versatility of the system. The paper's approach also lays the groundwork for extending pose estimation methodologies to handle multiple objects and more complex scenes.

In conclusion, HybridPose exemplifies a significant advancement in the field of computer vision, offering practical implementations for robotics and augmented reality while opening avenues for future explorations in hybrid representation techniques.