Learning shared template representation with augmented feature for multi-object pose estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deep learning-based template matching methods for pose estimation advance via metric or reconstruction learning.
• Current methods use object-specific template libraries, increasing training complexity and memory usage for multi-object tasks.
• These methods struggle with training-test distribution shifts and occlusions, impacting accuracy.

Purpose of the Study:

• To propose a shared template representation learning method with augmented semantic features for improved pose estimation.
• To address the limitations of object-specific template libraries and enhance generalization performance.
• To achieve superior matching accuracy and reduce memory costs in multi-object pose estimation.

Main Methods:

• Learns representations using concurrent metric and reconstruction learning as similarity constraints.
• Augments network response to objects via semantic feature constraints for better generalization.
• Utilizes rotation matrices as templates for codebook construction, decoupling object categories and templates.

Main Results:

• Achieves excellent matching accuracy compared to methods using rendered images.
• Demonstrates superior matching accuracy on Linemod, Linemod-Occluded, and TLESS datasets.
• Exhibits robustness on a collected aircraft dataset, validating efficacy.

Conclusions:

• The proposed shared template learning method effectively improves 3D object pose estimation accuracy.
• It offers a more efficient solution for multi-object tasks by maintaining a single shared codebook.
• The method shows significant potential for real-world applications requiring robust pose estimation, particularly with occlusions.