Few-shot Class-incremental Learning for Classification and Object Detection: A Survey

TL;DR

This survey systematically analyzes Few-shot Class-Incremental Learning (FSCIL), outlining its problem setting, core challenges, and a unified taxonomy for classification and object detection tasks. It distinguishes FSCIL from related paradigms (CIL, TIL, FSL), and groups methods into data-based, structure-based, and optimization-based categories, including detailed discussions of data replay, pseudo-scenarios, dynamic architectures, and meta-learning strategies. The work aggregates benchmark datasets and evaluation metrics, presents performance comparisons, and highlights gaps such as dataset scarcity and inconsistent protocols, while outlining forward-looking directions, practical settings, and safety considerations. Overall, the paper provides a structured foundation for advancing FSCIL research and guiding future deployments in dynamic, data-scarce environments.

Abstract

Few-shot Class-Incremental Learning (FSCIL) presents a unique challenge in Machine Learning (ML), as it necessitates the Incremental Learning (IL) of new classes from sparsely labeled training samples without forgetting previous knowledge. While this field has seen recent progress, it remains an active exploration area. This paper aims to provide a comprehensive and systematic review of FSCIL. In our in-depth examination, we delve into various facets of FSCIL, encompassing the problem definition, the discussion of the primary challenges of unreliable empirical risk minimization and the stability-plasticity dilemma, general schemes, and relevant problems of IL and Few-shot Learning (FSL). Besides, we offer an overview of benchmark datasets and evaluation metrics. Furthermore, we introduce the Few-shot Class-incremental Classification (FSCIC) methods from data-based, structure-based, and optimization-based approaches and the Few-shot Class-incremental Object Detection (FSCIOD) methods from anchor-free and anchor-based approaches. Beyond these, we present several promising research directions within FSCIL that merit further investigation.

Figures (9)

• Figure 1: IL publications from 2016 to 2023. It is observed that CIL research has become predominant in the field of IL over time, due to its practical value. Concurrently, FSCIL shows a steady rise, mirroring the growing requirement of CIL with limited data.
• Figure 2: A chronological overview of some representative FSCIL methods. FSCIL was first carried out by TOPIC tao2020few. CEC zhang2021few was widely used as a base for subsequent studies. SaKD cheraghian2021semantic integrated semantic word vectors into FSCIL, offering a reference for applying language-image models in the future. LCwoF kukleva2021generalized and ERDFR liu2022few proposed distinct Data Replay (DR) strategies. F2M shi2021overcoming introduced a novel approach by constraining optimization within flat local minima. FSLL+SS mazumder2021few introduced the semi-supervised features to FSCIL for the first time. MgSvF zhao2021mgsvf analyzed and utilized different frequency components to balance the old and new knowledge. FACT zhou2022forward introduced a fresh perspective by advocating forward compatibility in FSCIL. C-FSCIL hersche2022constrained pre-defined classifiers to guide model optimization. DSN yang2022dynamic offered a novel dynamic structure for FSCIL. LIMIT zhou2022few proposed a representative meta-learning paradigm for FSCIL. CLOM zou2022margin pointed out the issue of class-level overfitting made by metric learning in FSCIL.
• Figure 3: The general settings of different IL tasks. Specifically, (a) shows the setting of FSCIL, (b) is the CIL, (c) represents the setting of TIL, and (d) illustrates the Domain-incremental Learning (DIL). FSCIL can be viewed as a subdomain of CIL, where the base session usually has sufficient training data, and the incremental sessions are formed in the $N-$way $K-$shot format. TIL differs from CIL because the session identity is known during model training and testing. In contrast, DIL maintains the same classification tasks, but the data across different sessions comes from different domains. Note that "session" may also be called "task" in other literature.
• Figure 4: The illustration of unreliable empirical risk minimization in FSCIL. (a) with sufficient training samples, the empirical risk minimization can approximate the best-expected risk minimization function. (b) when the training samples are insufficient, the best empirical risk minimization function is often a poor approximation to the best-expected risk minimization function.
• Figure 5: The illustration of stability-plasticity dilemma in FSCIL. (a) and (b) are two consecutive sessions. Darker areas indicate optimal loss values. $\boldsymbol{\theta^p}$ performs well in session $p$ but poorly in $q$. Optimizing $\boldsymbol{\theta^p}$ to $\boldsymbol{\theta^q}$ on session $q$ diminishes its performance on session $p$. Yet, directing optimization towards $\boldsymbol{\theta^\star}$ ensures good results on both sessions.