A Study on Systematic Improvement of Transformer Models for Object Pose Estimation.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Transformer architectures, initially for NLP, are versatile for generative models.
• Object pose estimation is crucial for robotic manipulation and computer vision tasks.
• Existing transformer models for pose estimation face high GPU memory demands, especially for multiple objects.

Purpose of the Study:

• To develop a transformer-based deep learning model for efficient object pose estimation.
• To address high GPU memory consumption and improve training efficiency for multi-object scenarios.
• To enhance the performance of multi-object pose estimation using novel normalization techniques.

Main Methods:

• Implemented an encoder-only transformer architecture adapted for object pose estimation.
• Optimized the transformer's attention layer and incorporated low-rank weight decomposition to reduce parameters and memory usage.
• Applied Group Query Attention (GQA) and Root Mean Square (RMS) normalization for performance enhancement.

Main Results:

• Reduced GPU memory usage to 2.5% of the baseline model through an extended matrix dimension.
• Decreased model weight parameters by 28% using low-rank weight decomposition in the attention layer.
• Achieved a 17% improvement in rotation training accuracy over the baseline model with GQA and RMS normalization.

Conclusions:

• The proposed transformer-based model offers a significant reduction in memory footprint for object pose estimation.
• The integration of low-rank decomposition and advanced normalization techniques enhances both efficiency and accuracy.
• This work provides a more scalable and effective solution for multi-object pose estimation in computer vision and robotics.