Peak response regularization for localization.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep convolutional neural networks often assume Gaussian feature response, which fails with interference, causing sub-peaks and model drift.
• Progressive interference from background noise and other targets degrades model performance in tasks like object tracking and pose detection.

Purpose of the Study:

• To propose a novel feature response regularization approach for sub-peak response suppression and peak response enforcement.
• To systematically address and mitigate the effects of progressive interference in deep learning models.
• To enhance the localization and representation capabilities of convolutional features.

Main Methods:

• Introduced Peak Response Regularization (PRR), a method to aggregate and align discriminative features.
• Converted local extremal responses in discrete feature space to continuous space extremal responses.
• Applied PRR to enforce peak response and suppress sub-peak responses in feature maps.

Main Results:

• PRR demonstrated improved performance across multiple computer vision tasks, including human pose detection, object detection, and object tracking.
• The approach effectively suppressed sub-peak responses and enforced the main peak response on the tracking response map.
• Experiments confirmed enhanced localization and representation capabilities of convolutional features with PRR.

Conclusions:

• Peak Response Regularization (PRR) is an effective method for handling progressive interference in deep learning models.
• PRR significantly improves performance in various image-based tasks with negligible computational overhead.
• The proposed method enhances the robustness and accuracy of deep convolutional neural networks in challenging environments.