Semi-Supervised Human Detection via Region Proposal Networks Aided by Verification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised learning offers a promising avenue for improving model performance with limited labeled data.
• Human detection remains a challenging task, especially in complex, real-world scenarios.

Purpose of the Study:

• To enhance semi-supervised human detection by effectively utilizing unlabeled data.
• To develop a robust framework that minimizes the impact of noisy pseudo-annotations.

Main Methods:

• Modification of the Region Proposal Network (RPN) for partially labeled datasets.
• Introduction of a verification module to filter false positive proposals.
• Implementation of a self-paced training strategy for progressive pseudo-annotation incorporation.

Main Results:

• The proposed framework significantly improves human detection performance in a semi-supervised setting.
• State-of-the-art results achieved on scene-specific human detection benchmarks.
• Experimental validation confirms the effectiveness of the verification module and self-paced training.

Conclusions:

• Leveraging unlabeled data through refined region proposal and self-paced training is effective for semi-supervised human detection.
• The developed method offers a practical solution for improving human detection systems with limited annotations.
• This approach advances the state-of-the-art in semi-supervised computer vision tasks.