From Discriminant to Complete: Reinforcement Searching-Agent Learning for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Weakly supervised object detection (WSOD) aims to identify complete object instances using only image-level labels.
• Current WSOD methods often rely on proposal selection, which tends to identify discriminative object parts instead of whole objects.

Purpose of the Study:

• To develop a novel region searching paradigm for WSOD.
• To address the limitation of existing methods in detecting complete object instances.

Main Methods:

• A deep reinforcement learning framework is employed to train a searching agent.
• The searching process is modeled as a Markov decision process.
• Pseudo-complete object regions and discriminative object parts are extracted to create training pairs for the agent.

Main Results:

• The proposed learning strategy effectively mimics the searching process to reveal complete object regions from discriminative parts.
• The method avoids learning difficulties associated with long action sequences in full image searching.
• Integration with existing WSOD methods significantly improves performance over state-of-the-art and baseline approaches.

Conclusions:

• The developed searching agent effectively enhances weakly supervised object detection.
• This approach offers a promising direction for improving the identification of complete object instances in computer vision.