Deep Learning-Based 6-DoF Object Pose Estimation Considering Synthetic Dataset.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Generating high-quality 6-Degree-of-Freedom (6-DoF) object pose estimation datasets is challenging.
• Domain gaps between synthetic and real data limit the accuracy and generalization of existing pose estimation methods.

Purpose of the Study:

• To propose a novel methodology for enhancing 6-DoF object pose estimation accuracy and generalization.
• To address the domain gap issue using improved datasets and deep learning techniques.

Main Methods:

• Utilized Blenderproc for high-quality synthetic data generation, processed with bilateral filtering to minimize domain gaps.
• Developed an attention-based mask region-based convolutional neural network (R-CNN) for improved detection accuracy and reduced computational cost.
• Introduced an improved feature pyramidal network (iFPN) with added bottom-up paths for enhanced feature extraction.
• Proposed a novel convolutional block attention module-convolutional denoising autoencoder (CBAM-CDAE) network incorporating channel and spatial attention mechanisms.
• Implemented pose refinement for accurate 6-DoF object pose determination.

Main Results:

• The proposed attention-based mask R-CNN significantly improved detection accuracy.
• The iFPN effectively extracted deeper image features.
• The CBAM-CDAE enhanced the autoencoder's feature extraction capabilities.
• Evaluations on T-LESS and LineMOD datasets demonstrated superior performance compared to existing models.

Conclusions:

• The proposed methodology effectively reduces the domain gap between synthetic and real data.
• The novel deep learning architecture achieves state-of-the-art accuracy in 6-DoF object pose estimation.
• This approach offers a promising solution for real-world applications requiring precise object pose information.