Multiple Instance Differentiation Learning for Active Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Active learning significantly advances image recognition but lacks systematic instance-level investigation for object detection.
• Object detection models require efficient data selection strategies, particularly when labeled datasets are small.

Purpose of the Study:

• To propose a novel method for instance-level active learning in object detection.
• To unify instance uncertainty and image uncertainty for more informative image selection.
• To establish a strong baseline for instance-level active learning in object detection.

Main Methods:

• Developed Multiple Instance Differentiation Learning (MIDL) comprising classifier prediction and multiple instance differentiation modules.
• Employed adversarial instance classifiers to estimate instance uncertainty in unlabeled data.
• Utilized a multiple instance learning approach to re-estimate image-instance uncertainty, unifying it with instance uncertainty via Bayesian theory.

Main Results:

• MIDL effectively unifies image and instance uncertainties within a Bayesian framework.
• Extensive experiments demonstrate MIDL's superior performance over state-of-the-art methods on standard object detection datasets.
• The proposed method shows significant improvements, especially in scenarios with limited labeled data.

Conclusions:

• MIDL provides a robust and effective approach for instance-level active learning in object detection.
• The method offers a significant advancement for improving the efficiency of object detection model training.
• MIDL establishes a new benchmark for active learning strategies in object detection tasks.