Multi-Scale Low-Discriminative Feature Reactivation for Weakly Supervised Object Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object localization (WSOL) faces challenges in preventing networks from focusing on limited discriminative object parts.
• Class Activation Mapping (CAM) based methods often use Adversarial Learning (AL), which can be difficult to control and may limit localization capabilities.

Purpose of the Study:

• To propose a novel framework, Multi-scale Low-Discriminative Feature Reactivation (mLDFR), for WSOL.
• To improve the network's ability to localize objects by reactivating low-discriminative parts.
• To enhance localization power without sacrificing classification accuracy and enable multi-instance localization.

Main Methods:

• The proposed mLDFR framework utilizes bottom-up continuous feature map recalibration.
• It incorporates multi-scale object category mapping to identify and reactivate underutilized object features.
• The framework is designed to be flexible and compatible with various Convolutional Neural Network (CNN) backbones.

Main Results:

• The mLDFR framework significantly improves localization power while maintaining classification accuracy.
• It enables multi-instance localization, a capability lacking in many AL-based frameworks.
• Achieved state-of-the-art results on ILSVRC2014 and CUB200-2011 datasets, demonstrating superior performance.

Conclusions:

• The mLDFR framework offers a robust and effective solution for WSOL.
• It overcomes limitations of AL-based methods by enhancing feature utilization and localization capabilities.
• The method demonstrates significant advancements in object localization accuracy and flexibility.