Combining Synthetic Images and Deep Active Learning: Data-Efficient Training of an Industrial Object Detection Model.

Area of Science:

• Computer Vision
• Machine Learning
• Data Augmentation

Background:

• Limited training data is a significant challenge in industrial deep learning.
• A domain gap exists between synthetic and real-world data, impacting model performance.
• Combining synthetic and real data sources generally yields superior results.

Purpose of the Study:

• To develop a data-efficient workflow for training industrial object detection models.
• To iteratively improve model performance using a combination of synthetic data and active learning.
• To address the sim-to-real domain gap in industrial computer vision applications.

Main Methods:

• Physics-based rendering was used to generate synthetic training images.
• Deep active learning was employed to iteratively refine the object detection model.
• A hybrid query strategy was implemented to select informative training samples.

Main Results:

• Synthetic images significantly boosted model performance, particularly with limited initial real-world data.
• The hybrid active learning strategy outperformed random sampling in selecting diverse and informative images.
• The workflow demonstrated improved model accuracy and efficiency.

Conclusions:

• A practical workflow for training and improving object detection models with minimal real-world data was established.
• The integration of synthetic data and active learning leads to data-efficient and cost-effective computer vision solutions.
• This approach is highly beneficial for industrial applications with data scarcity.