Augmentation Invariant and Instance Spreading Feature for Softmax Embedding.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep embedding learning is crucial for discriminative feature representations.
• Unsupervised learning aims to create these representations without category labels.
• Challenges include mining supervision from fine-grained classes and generalizing to unseen categories.

Purpose of the Study:

• To develop an unsupervised embedding learning method for fine-grained visual categorization.
• To address challenges in positive supervision mining and generalization to unseen categories.
• To approximate supervised learning properties using instance-wise supervision.

Main Methods:

• Introduced a data augmentation invariant and instance spreading feature using instance-wise supervision.
• Designed two domain-agnostic augmentation strategies to enhance feature space supervision.
• Proposed a novel instance-wise softmax embedding optimizing over augmented instance features with binary discrimination.

Main Results:

• The proposed unsupervised embedding learning significantly accelerates learning speed and improves accuracy.
• Achieved high performance on both seen and unseen testing categories.
• Demonstrated effectiveness even without a pre-trained network on fine-grained samples.
• A variant using category-wise supervision achieved competitive results against state-of-the-art methods.

Conclusions:

• The novel instance-wise softmax embedding offers an effective approach to unsupervised embedding learning for fine-grained categories.
• The method generalizes well to unseen categories and performs robustly without pre-training.
• The approach provides a strong alternative to supervised methods, even with category-wise supervision.