Min-Entropy Latent Model for Weakly Supervised Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised object detection faces challenges due to inconsistent supervision, leading to random object locations and ambiguous detectors.
• Existing methods struggle to simultaneously learn object locations and detectors effectively under weak supervision.

Purpose of the Study:

• To propose a novel min-entropy latent model (MELM) for addressing the challenges in weakly supervised object detection.
• To reduce the variance of learned instances and alleviate ambiguity in object detectors.

Main Methods:

• The proposed MELM utilizes min-entropy as both a model for learning object locations and a metric for measuring localization randomness.
• MELM is architecturally decomposed into proposal clique partition, object clique discovery, and object localization components.
• A recurrent learning algorithm employing continuation optimization is used to address the non-convexity inherent in MELM optimization.

Main Results:

• MELM demonstrated significant performance improvements in weakly supervised object detection.
• The model also achieved notable gains in weakly supervised object localization and image classification tasks.
• Experimental results show MELM outperforming current state-of-the-art approaches.

Conclusions:

• The proposed min-entropy latent model (MELM) effectively addresses key challenges in weakly supervised object detection.
• MELM offers a principled approach to reduce localization randomness and detector ambiguity.
• The method shows superior performance compared to existing state-of-the-art techniques across multiple related tasks.