Semi-Supervised Domain Adaptation for Holistic Counting under Label Gap.

Area of Science:

• Computer Science
• Machine Learning
• Computer Vision

Background:

• Holistic regression tasks, predicting continuous values from images, often face domain shift and label gaps in real-world datasets.
• Existing domain adaptation methods primarily focus on classification tasks, leaving a gap for regression.

Purpose of the Study:

• To propose a novel, unified framework for semi-supervised domain adaptation in holistic regression.
• To address both covariate shift and label gaps simultaneously.

Main Methods:

• Utilized a Generative Adversarial Network (GAN) to minimize covariate shift.
• Employed label normalization to mitigate label gaps.
• Developed a stopping criterion combining Maximum Mean Discrepancy and GAN Global Optimality to prevent overfitting.
• Incorporated a small set of target domain annotations to restore the original label range.

Main Results:

• Demonstrated significant outperformance over state-of-the-art methods across three diverse datasets.
• Achieved drastic reductions in Mean Squared Error (MSE) for cell counting (759 to 5.62) and pedestrian detection (131 to 1.47).
• Showcased substantial MSE improvements in plant biology tasks, including leaf counting (intra-species: 2.36 to 0.88; inter-species: 1.48 to 0.6).

Conclusions:

• The proposed unified framework effectively handles both covariate shift and label gaps in semi-supervised domain adaptation for holistic regression.
• The method offers a robust solution for real-world regression problems where data discrepancies are common.
• Validated through extensive experiments, the approach represents a significant advancement in domain adaptation for regression tasks.