Instance-Level Few-Shot Learning With Class Hierarchy Mining.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning addresses challenges with limited training data for new classes.
• Instance-level few-shot learning often overlooks inter-category relationships.
• Effective feature utilization from base classes is crucial for novel class recognition.

Purpose of the Study:

• To exploit hierarchical information for improved few-shot instance segmentation (FSIS).
• To leverage discriminative features from base classes for novel object classification.
• To introduce a novel superclass approach for enhancing FSIS performance.

Main Methods:

• Proposed a superclass approach to automatically create a class hierarchy for FSIS.
• Designed the Soft Multiple Superclass (SMS) framework to extract relevant features within superclasses.
• Applied label refinement to train the hierarchy-based detector and describe class associations.

Main Results:

• The proposed method demonstrates significant effectiveness on FSIS benchmarks.
• Leveraging hierarchical information improves the classification of novel objects with scarce data.
• The SMS framework successfully extracts relevant features by grouping classes into superclasses.

Conclusions:

• The novel superclass approach and SMS framework offer an effective solution for FSIS.
• Exploiting category hierarchies is beneficial for few-shot learning tasks.
• The method provides a robust way to utilize base class features for novel class recognition.