HiFormer: Hierarchical Transformer With Box-Packed Positional Encoding for 3D Part Assembly.

Area of Science:

• Computer Graphics
• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Estimating the 6-DoF posture of parts is crucial for assembly-based modeling.
• Automating tasks like IKEA furniture assembly requires advanced 3D part understanding.

Purpose of the Study:

• To present HiFormer, a novel Hierarchical Transformer with Box-packed Positional Encoding for highly automatic 3D part assembly.
• To address overfitting in Transformer-based 3D point cloud feature learning.
• To effectively model intragroup and intergroup part relationships.
• To compute and integrate positional encoding for diverse part geometries in assembly.

Main Methods:

• Utilized a multi-task 3D Swin Transformer with a two-stage training strategy for feature extraction.
• Developed a hierarchical Transformer to capture part relationships at multiple levels (flattening, intragroup, intergroup).
• Introduced innovative box-packed positional encoding incorporating relative box positions into the Transformer.

Main Results:

• HiFormer outperformed the state-of-the-art PWH-MP model on the PartNet benchmark (Chair, Table, Lamp categories).
• Achieved average improvements of 2.84% in Part Accuracy (PA) and 3.72% in Connection Accuracy (CA) for diversity modeling.
• Demonstrated average improvements of 3.55% in PA and 3.21% in CA for deterministic modeling.

Conclusions:

• The proposed HiFormer method effectively enhances automatic 3D part assembly.
• The novel hierarchical structure and box-packed positional encoding significantly improve accuracy in part posture estimation.
• HiFormer offers a robust solution for complex assembly tasks in robotics and computer vision.