Instance-Level Contrastive Learning for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object detection (WSOD) reduces labeling costs by using image-level annotations.
• Existing WSOD methods often overlook instance correlations across different images.
• Learning robust instance representations is crucial for improving WSOD performance.

Purpose of the Study:

• To propose an instance-level contrastive learning (ICL) framework for WSOD.
• To address the limitation of ignoring instance correlations in existing WSOD methods.
• To enhance the mining of reliable object instance representations by leveraging inter-image relationships.

Main Methods:

• Proposed an instance-level contrastive learning (ICL) framework.
• Introduced an instance-diverse memory updating (IMU) algorithm for a memory bank with multiple representation vectors per class.
• Developed memory-aware instance mining (MIM) and memory-aware proposal sampling (MPS) algorithms.

Main Results:

• The proposed ICL framework significantly improved WSOD performance on PASCAL VOC2007 and VOC2012 datasets.
• Achieved 14.2% mAP and 13.4% CorLoc gains on PASCAL VOC2007.
• Achieved 12.2% mAP and 8.3% CorLoc gains on PASCAL VOC2012 compared to the baseline.

Conclusions:

• The proposed ICL framework effectively mines reliable instance representations by considering inter-image correlations.
• The IMU, MIM, and MPS algorithms contribute to improved foreground representation and balanced positive-negative sample learning.
• The method demonstrates substantial performance gains, highlighting its effectiveness for WSOD.