Generative Data Augmentation for ArUco-Free RGB-Based 6-DoF Object Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Data-driven methods are crucial for industrial 6-Degrees-of-Freedom (6-DoF) object pose estimation.
• Benchmark datasets can contain biases, like ArUco markers, that compromise model reliability.
• These markers are often absent in real industrial settings but are exploited by neural networks.

Purpose of the Study:

• Investigate shortcut bias introduced by ArUco markers in the Linemod dataset.
• Assess the impact of these biases on 6-DoF object pose estimation performance.
• Propose a mitigation strategy using generative AI for more reliable industrial applications.

Main Methods:

• Saliency map analysis to visualize model attention on ArUco markers.
• Data augmentation using generative AI to remove markers and synthesize realistic backgrounds.
• Experimental evaluation of models trained on original vs. augmented datasets in ArUco-free environments.

Main Results:

• Saliency maps confirmed significant model attention on ArUco markers, indicating shortcut learning.
• Models trained on original Linemod showed performance degradation in ArUco-free settings.
• Training with generative AI-augmented data improved model robustness and generalization.

Conclusions:

• Background-induced biases in pose estimation benchmarks significantly impact model performance.
• Generative AI-based augmentation offers a practical approach to mitigate these biases.
• The proposed method enhances reliability for industrial 6-DoF pose estimation systems, though further improvements are possible.