VILOD: Combining Visual Interactive Labeling With Active Learning for Object Detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• High-quality annotated datasets are crucial for training robust object detection (OD) models.
• Current methods for dataset creation often face limitations in scale and quality.
• Active Learning (AL) offers a promising approach to optimize annotation efforts.

Purpose of the Study:

• To introduce VILOD, a Visual Interactive Labeling tool designed for efficient OD annotation.
• To integrate Active Learning (AL) with interactive visualizations for a Human-in-the-Loop (HITL) workflow.
• To enable expert users to implement strategic, visually guided labeling strategies.

Main Methods:

• Development of VILOD, a novel interactive labeling tool.
• Integration of AL with interactive visualizations for transparent and steerable annotation.
• Comparative case studies evaluating visually guided strategies against automated AL baselines.

Main Results:

• Visually guided labeling strategies, particularly a balanced human-guided approach, outperformed the automated AL baseline.
• VILOD's visual cues facilitated synthesis of data structure and model uncertainty information.
• The human-guided strategy achieved the highest overall OD model performance.

Conclusions:

• Interactive, visually guided annotation significantly enhances the efficiency and effectiveness of OD dataset creation.
• VILOD empowers expert users to develop superior labeling strategies.
• Human-in-the-Loop (HITL) workflows augmented with visual intelligence are key for advancing OD model training.