High-Quality Proposals for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly Supervised Object Detection (WSOD) faces challenges in generating and selecting accurate object proposals.
• Existing methods often struggle with generating proposals that fully enclose objects or selecting discriminative negative proposals.

Purpose of the Study:

• To address the key challenges in WSOD: proposal generation and proposal selection.
• To develop a unified framework that improves the quality and effectiveness of object proposals for WSOD.

Main Methods:

• Proposal generation combines selective search with Gradient-weighted Class Activation Mapping (Grad-CAM) for higher Intersection-Over-Union (IOU) with ground truth.
• Proposal selection focuses on confident positive proposals and class-specific hard negatives, up-weighting their losses for more effective training.
• The unified approach is integrated into the Online Instance Classifier Refinement (OICR) framework.

Main Results:

• Significant improvements in mean Average Precision (mAP) and Correct Localization (CorLoc) on PASCAL VOC 2007, VOC 2012, and MS COCO datasets.
• Demonstrated substantial gains over the baseline OICR method across all tested datasets.
• Achieved state-of-the-art performance compared to existing WSOD methods.

Conclusions:

• The proposed methods for proposal generation and selection are effective and broadly applicable to WSOD.
• The unified approach significantly enhances the performance of WSOD systems.
• This work sets a new state-of-the-art in weakly supervised object detection.