Mixed Supervised Object Detection with Robust Objectness Transfer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised detection (WSD) struggles with new object categories due to limited labeled data.
• Existing mixed supervised detection (MSD) methods often fail to generalize by directly transferring detectors.
• Distractors, like object parts, in weakly labeled images pose a significant challenge.

Purpose of the Study:

• To propose a robust objectness transfer approach for mixed supervised detection (MSD).
• To improve the detection of new object categories by leveraging existing fully labeled categories.
• To enhance the ability of models to reject distractors in weakly labeled images.

Main Methods:

• Learning domain-invariant objectness knowledge from fully labeled categories.
• Modeling knowledge using invariant features robust to distribution discrepancies.
• Utilizing multiple instance learning (MIL) guided by objectness knowledge to model objects and distractors.

Main Results:

• The proposed objectness transfer approach outperforms existing MSD methods.
• Achieved state-of-the-art results on challenging datasets like ILSVRC2013 and PASCAL VOC.
• Demonstrated improved rejection of distractors in weakly labeled images.

Conclusions:

• The robust objectness transfer method effectively improves MSD performance.
• Domain-invariant knowledge learning is crucial for generalizing to new object categories.
• The approach offers a more reasonable and robust solution for mixed supervised detection tasks.