Contrastive Proposal Extension With LSTM Network for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Weakly supervised object detection (WSOD) uses image-level labels to reduce annotation costs.
• Existing WSOD methods, often based on Multiple Instance Learning (MIL), struggle with incomplete object region detection.

Purpose of the Study:

• To address the insufficient integrity issue in MIL-based WSOD methods.
• To propose a novel WSOD strategy that optimizes initial object proposals for better completeness.

Main Methods:

• Introduced Contrastive Proposal Extension (CPE) with Multiple Directional CPEs (D-CPEs).
• Utilized LSTM-based encoders and dual-stream decoders to compare initial and extended proposals.
• Extended proposal boundaries sequentially and analyzed feature semantics for proposal integrity optimization.

Main Results:

• The proposed method significantly improved the integrity of detected object regions.
• Achieved state-of-the-art performance on PASCAL VOC 2007, VOC 2012, and MS-COCO datasets.
• Demonstrated effective suppression of inaccurate proposals and enhancement of correct ones.

Conclusions:

• Contrastive Proposal Extension (CPE) effectively enhances object region completeness in WSOD.
• The integration of LSTM encoders and dual-stream decoders contributes to improved WSOD performance.
• The method offers a promising direction for more accurate and cost-effective object detection.