Friend or Foe: Fine-Grained Categorization With Weak Supervision

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multi-instance learning (MIL) is a standard technique for weakly supervised problems.
• Fine-grained visual categorization presents extreme weakly supervised challenges due to high intra-class variance and subtle inter-class differences.

Purpose of the Study:

• To investigate the applicability of MIL in extreme weakly supervised learning for fine-grained visual categorization.
• To propose a generalized MIL framework addressing the challenges of subtle category distinctions.

Main Methods:

• A novel multi-task co-localization algorithm is proposed to leverage relationships among fine-grained categories.
• This algorithm serves as an effective initialization for the non-convex MIL objective.
• Localization results facilitate object-level domain-specific fine-tuning of deep neural networks.

Main Results:

• The proposed method demonstrates effectiveness on three fine-grained datasets.
• Exploiting inter-class relationships significantly enhances weakly supervised fine-grained recognition.
• Object-level fine-tuning boosts overall performance.

Conclusions:

• The generalized MIL framework with multi-task co-localization is effective for fine-grained visual categorization.
• Leveraging inter-class relationships is crucial for improving weakly supervised recognition in fine-grained tasks.
• The proposed method offers a significant advancement in handling extreme weakly supervised learning scenarios.