Toward Real-World Category-Level Articulation Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Current Category-level Articulation Pose Estimation (CAPE) methods are limited to single instances and fixed kinematic structures.
• Real-world scenarios involve multiple articulated objects with varying kinematic structures in a single observation.

Purpose of the Study:

• To address limitations in existing CAPE methods by proposing a new task setting, CAPE-Real (CAPER), for estimating part-level 6D pose of multiple articulated objects with unknown kinematic structures from RGB-D images.
• To develop a robust framework capable of handling multi-instance scenarios and unknown kinematic structures.

Main Methods:

• Introduced the CAPE-Real (CAPER) task setting, allowing varied kinematic structures and multiple instances.
• Developed an articulated model repository (ReArt-48) and a dataset generation pipeline (FAOM and SAMERT).
• Created large-scale mixed reality (ReArtMix) and real-world (ReArtVal) datasets.
• Proposed a framework utilizing RGB-D input, object detection, instance segmentation, an Articulation Parsing Network, and a Pair Articulation Pose Estimation module.

Main Results:

• The proposed method effectively estimates part-level pose for multiple instances in a single forward pass.
• Achieved strong performance on the CAPER task, as well as on existing CAPE and robot arm pose estimation benchmarks.
• Demonstrated the capability to handle unknown kinematic structures and multi-instance scenarios.

Conclusions:

• The developed framework and CAPER task setting provide a strong baseline for future research in articulated object pose estimation.
• The publicly available datasets and code will facilitate further advancements in the field.
• The approach shows promise for real-world applications requiring accurate 6D pose estimation of articulated objects.