Weakly Supervised Object Localization with Multi-Fold Multiple Instance Learning.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object category localization is a difficult computer vision task.
• Standard methods require time-consuming bounding box annotations.
• Weakly supervised learning uses only image-level labels, avoiding manual annotation.

Purpose of the Study:

• To develop a weakly supervised learning approach for object localization.
• To improve object detection accuracy without precise bounding box supervision.
• To address the challenge of erroneous location inference in training.

Main Methods:

• A multiple-instance learning (MIL) framework is employed.
• A multi-fold MIL procedure is introduced to prevent premature convergence on incorrect locations.
• A window refinement method incorporating an objectness prior is proposed.

Main Results:

• The multi-fold MIL procedure enhances robustness, especially with high-dimensional features (e.g., Fisher vectors, CNN features).
• The window refinement method improves localization accuracy.
• Experimental evaluation on the PASCAL VOC 2007 dataset demonstrates the approach's effectiveness.

Conclusions:

• The proposed weakly supervised object localization method is effective and efficient.
• The multi-fold MIL and window refinement techniques significantly improve localization performance.
• This work offers a viable alternative to fully supervised methods for object localization.