Pyramidal Multiple Instance Detection Network With Mask Guided Self-Correction for Weakly Supervised Object...

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Weakly supervised object detection (WSOD) utilizes image-level annotations, reducing data labeling costs.
• Multiple Instance Detection Networks (MIDNs) are a common WSOD approach, integrating multiple instance learning with deep convolutional neural networks.
• A key challenge in MIDNs is their tendency to converge to local discriminative regions, hindering complete object localization.

Purpose of the Study:

• To address the local optimum issue in MIDNs for improved weakly supervised object detection.
• To propose a novel pyramidal MIDN (P-MIDN) architecture.
• To enhance the training of online instance classifier refinement (OICR) frameworks using P-MIDN.

Main Methods:

• Introduced a pyramidal MIDN (P-MIDN) consisting of a sequence of MIDNs.
• Implemented a proposal removal mechanism where earlier MIDNs filter proposals for subsequent ones, preventing focus on limited discriminative regions.
• Integrated P-MIDN with an online instance classifier refinement (OICR) framework.
• Developed a mask guided self-correction (MGSC) method for generating high-quality pseudo ground-truths to train the OICR.

Main Results:

• The proposed P-MIDN effectively mitigates the local optimum problem by enabling focus on more complete object proposals.
• Integration with OICR and MGSC resulted in a robust framework for weakly supervised object detection.
• Achieved state-of-the-art performance on several benchmark datasets including PASCAL VOC (2007, 2010, 2012), ILSVRC 2013 DET, and MS-COCO.

Conclusions:

• The pyramidal MIDN effectively overcomes limitations of standard MIDNs in weakly supervised object detection.
• The proposed method demonstrates superior performance and robustness across diverse and challenging benchmarks.
• This work offers a significant advancement in weakly supervised object detection techniques.