Weakly Supervised Object Detection Using Proposal- and Semantic-Level Relationships.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Weakly supervised object detection (WSOD) is a challenging computer vision problem.
• Existing WSOD methods often overlook contextual information, relying solely on visual appearance.

Purpose of the Study:

• To develop a novel deep multiple instance reasoning framework for WSOD.
• To leverage both proposal-level and semantic-level context to improve WSOD performance.

Main Methods:

• Introduced two levels of context: proposal-level (spatial relationships) and semantic-level (category co-occurrence).
• Developed a framework incorporating graph convolutional networks (GCNs) for object location and multi-label reasoning.
• Integrated these components into an end-to-end training procedure with a CNN backbone.

Main Results:

• The proposed framework demonstrated superior performance compared to state-of-the-art methods and baselines.
• Experiments were conducted on widely used PASCAL VOC and MS COCO benchmarks.
• The approach effectively utilizes contextual information for improved object detection.

Conclusions:

• The novel deep multiple instance reasoning framework significantly advances WSOD.
• Incorporating contextual reasoning alongside visual appearance is crucial for tackling the ill-posed nature of WSOD.
• The method shows strong potential for real-world object detection applications.