Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Few-shot Class-Incremental Learning (FSCIL) addresses the challenge of learning new classes from limited data while retaining knowledge of previously learned classes.
• This area of machine learning is crucial for developing adaptive and continuously learning systems.

Purpose of the Study:

• To provide a comprehensive and systematic review of the Few-shot Class-Incremental Learning (FSCIL) field.
• To consolidate understanding of FSCIL's core challenges, methodologies, and future research avenues.

Main Methods:

• Systematic literature review of FSCIL.
• Categorization of FSCIL methods into data-based, structure-based, and optimization-based approaches for classification (FSCIC).
• Categorization of FSCIL methods into anchor-free and anchor-based approaches for object detection (FSCIOD).

Main Results:

• Detailed examination of FSCIL problem definition, challenges (unreliable empirical risk minimization, stability-plasticity dilemma), and general schemes.
• Overview of benchmark datasets and evaluation metrics for FSCIL.
• Introduction and categorization of existing FSCIC and FSCIOD methods.

Conclusions:

• FSCIL remains an active research area with significant challenges and evolving methodologies.
• Identified several promising future research directions within FSCIL warranting further investigation.