Small Object Localization with 90% Annotation Reduction by Positive-Unlabeled Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Small object localization is difficult due to poor visual appearance and noisy data.
• Current methods often rely on extensive annotations, increasing training costs.
• Human learning demonstrates efficient skill acquisition with limited examples.

Purpose of the Study:

• To develop a novel approach for small object localization using positive-unlabeled (PU) learning.
• To enable accurate localization with significantly reduced annotation effort.
• To simulate human-like learning from partial annotation data.

Main Methods:

• Proposed a novel positive-unlabeled (PU) learning framework.
• Utilized partial point annotations for training the localization model.
• Evaluated the approach on diverse small object datasets (single cell, animal/insect, crowds).

Main Results:

• Achieved strong localization performance with an F1 score exceeding 0.75.
• Demonstrated effectiveness even with less than 10% of total point annotations.
• Validated the approach across multiple challenging small object scenarios.

Conclusions:

• The proposed PU learning method offers an effective solution for small object localization.
• Significantly reduces annotation costs, making it suitable for data-scarce applications.
• Enables low-annotation-cost analysis, particularly for single-cell studies in microfluidics.