Pov9D: Point Cloud-Based Open-Vocabulary 9D Object Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Traditional object pose estimation relies heavily on RGB data and appearance cues.
• Textureless or feature-poor objects present significant challenges for existing methods, leading to pose ambiguities.
• Open-vocabulary settings require models to generalize to unseen object categories.

Purpose of the Study:

• To introduce Pov9D, a novel point cloud-based framework for robust open-vocabulary object pose estimation.
• To leverage 3D geometric information and semantic textual descriptions for improved pose accuracy.
• To enable accurate pose estimation for objects with limited or no texture.

Main Methods:

• Utilizes 3D point clouds as primary input, focusing on geometric structures.
• Integrates category-level textual descriptions to guide the pose estimation process.
• Introduces a text-conditioned shape prior generator to predict normalized object shapes from point clouds and text.

Main Results:

• Pov9D achieves state-of-the-art performance on the OO3D-9D benchmark.
• Demonstrates significant improvements in Absolute IoU@50 (+7.2%) and Relative 10° 10 cm (+27.2%) compared to OV9D.
• Effectively predicts object translation, rotation, and size, even for previously unseen categories.

Conclusions:

• Point cloud-based approaches combined with textual guidance offer a robust solution for open-vocabulary object pose estimation.
• Pov9D's text-conditioned shape prior enables accurate pose prediction by providing a consistent geometric reference.
• The framework shows strong generalization capabilities, addressing limitations of appearance-based methods.