Penalizing the Hard Example But Not Too Much: A Strong Baseline for Fine-Grained Visual Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Fine-grained visual classification (FGVC) faces challenges with model generalization due to severe overfitting.
• Existing FGVC methods often overfit hard training samples, failing to generalize to unseen hard examples in test sets.

Purpose of the Study:

• To introduce a novel strategy, moderate hard example modulation (MHEM), to address overfitting in FGVC.
• To enhance model generalization and discrimination capabilities for improved performance on unseen data.

Main Methods:

• Propose a moderate hard example modulation (MHEM) strategy to modulate hard examples effectively.
• Formulate a general form of a modulated loss function based on three conditions.
• Instantiate the loss function to create a strong baseline for FGVC.

Main Results:

• The proposed MHEM strategy encourages models to avoid overfitting hard examples, leading to better generalization.
• A naive backbone integrated with MHEM achieves performance comparable to recent FGVC methods.
• The MHEM baseline can be incorporated into existing methods, enhancing their discriminative power.

Conclusions:

• Moderate hard example modulation (MHEM) offers a promising approach to improve generalization and discrimination in FGVC.
• The proposed method consistently improves performance on standard FGVC datasets like CUB-200-2011, Stanford Cars, and FGVC-Aircraft.
• MHEM has the potential to inspire future research in fine-grained visual recognition.