Automatic Bounding Box Annotation with Small Training Datasets for Industrial Manufacturing.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Object detection is crucial for human-robot collaboration and Industry 5.0.
• Deep learning advancements have improved object detection quality.
• Automatic training data generation for new objects is a bottleneck in industrial manufacturing.

Purpose of the Study:

• To adapt state-of-the-art object detection methods for automatic bounding box annotation.
• To enable models to learn new objects in changing environments with minimal data.
• To address limitations in current industrial applications of object detection.

Main Methods:

• Adapted Faster R-CNN and Scaled-YOLOv4-p5 architectures.
• Focused on a use case with a homogeneous background and human-provided object labels.
• Trained models to distinguish unknown objects from complex backgrounds using limited data.

Main Results:

• Both adapted architectures successfully distinguished unknown objects from homogeneous backgrounds.
• The proposed method requires significantly less training data compared to other state-of-the-art approaches.
• Outperformed the LOST (transformer-based object discovery) method on a fruits dataset.

Conclusions:

• Adapted object detection models can effectively perform automatic bounding box annotation with minimal training data.
• The proposed method removes the need for human verification, predefined classes, or large annotated datasets.
• This approach significantly advances the practical application of object detection in industrial settings.