Increasing the Robustness of Deep Learning Models for Object Segmentation: A Framework for Blending Automatically...

Area of Science:

• Robotics
• Machine Learning
• Computer Vision

Background:

• Deep learning (DL) and transfer learning are crucial for modern robotics.
• Robustness against environmental variations (e.g., illumination) is essential for fine-tuned models.
• Limited research exists on using synthetic data for fine-tuning DL models.

Purpose of the Study:

• To propose methods for automatic generation of annotated image datasets for object segmentation (real and synthetic).
• To introduce a novel domain adaptation approach, Filling the Reality Gap (FTRG), for blending real and synthetic data.
• To evaluate the effectiveness of synthetic data for fine-tuning in transfer learning and continual learning.

Main Methods:

• Developed automated methods for generating annotated real-world and synthetic image datasets.
• Introduced the Filling the Reality Gap (FTRG) domain adaptation technique.
• Conducted experiments on a robotic application, comparing FTRG with domain randomization and photorealistic synthetic images.

Main Results:

• FTRG demonstrated superior performance in creating robust models compared to existing domain adaptation techniques.
• Fine-tuning with synthetic data, using proposed methods and FTRG, yielded better results than using only real-world data.
• Synthetic data for fine-tuning in transfer learning and continual learning proved beneficial.

Conclusions:

• Automated synthetic data generation and FTRG are effective for domain adaptation in robotics.
• Fine-tuning with synthetic data enhances the robustness and performance of deep learning models in robotic applications.
• The proposed methods offer practical solutions for overcoming limitations in synthetic data generation for fine-tuning.