PML-LocNet: Improving Object Localization with Prior-induced Multi-view Learning Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly Supervised Object Localization (WSOL) relies solely on image-level labels, making accurate object location inference challenging.
• Existing WSOL methods often struggle with sensitivity to initial conditions and overfitting due to alternating optimization strategies.
• The accurate localization of objects is a critical step for many computer vision tasks.

Purpose of the Study:

• To develop a novel model, PML-LocNet, that enhances object localization accuracy in WSOL problems.
• To address the limitations of previous methods by improving robustness and reducing overfitting.
• To leverage both view and sample diversity for more reliable object localization.

Main Methods:

• Introduced the Prior-induced Multi-view Learning Localization Network (PML-LocNet).
• Employed a two-phase multi-view learning strategy to exploit view diversity, enhancing feature complementarity and localization consensus.
• Utilized coarse-to-fine priors at image and instance levels to harness sample diversity, prioritizing reliable samples for robust network learning.

Main Results:

• PML-LocNet demonstrated significant improvements in localization accuracy.
• Achieved 69.3% CorLoc and 50.4% mAP on the PASCAL VOC 2007 dataset.
• Outperformed existing state-of-the-art methods by a considerable margin.

Conclusions:

• PML-LocNet effectively improves object localization in WSOL tasks by integrating multi-view and sample diversity.
• The model is robust and can be easily integrated with existing WSOL frameworks.
• Experimental results validate the superior performance of PML-LocNet over current approaches.