Coarse-to-Fine Adaptive People Detection for Video Sequences by Maximizing Mutual Information ^†

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• People detectors struggle with unseen data due to variations in viewpoints, poses, and occlusions.
• Existing methods often require manual labeling for new environments, limiting adaptability.

Purpose of the Study:

• To develop a coarse-to-fine framework for real-time adaptation of people detectors.
• To improve the robustness and accuracy of people detection on diverse, unannotated data.

Main Methods:

• Proposes a framework that adapts people detectors frame-by-frame during runtime.
• Utilizes mutual information between multiple detectors to estimate similarities and dissimilarities.
• Identifies representative frames globally and optimizes detection thresholds locally for each detector.

Main Results:

• The proposed approach adapts detectors without requiring additional manually labeled ground truth.
• Achieves superior performance compared to state-of-the-art detectors with fixed, pre-determined thresholds.
• Demonstrates effectiveness in handling variations in viewpoints, motion, poses, backgrounds, occlusions, and people sizes.

Conclusions:

• The coarse-to-fine framework offers an effective solution for adapting people detectors to unseen data.
• Runtime adaptation using mutual information significantly enhances detection performance without retraining.
• This method provides a robust and efficient approach for real-world people detection applications.