Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey
Open accessThe DOI matches its Crossref record; 123 of the 123 references it cites were found in a registry or an index; 54 of the 100 sources its claims cite were checked for support.
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.
Overview
CoChat Summary
TL;DR
This survey systematically reviews few-shot class-incremental learning (FSCIL) for classification and object detection, covering definitions, challenges, datasets, metrics, methods, and future directions. It organizes FSCIC methods into data-, structure-, and optimization-based approaches and FSCIOD methods into anchor-free and anchor-based approaches.
Key findings
- FSCIL’s core challenges are unreliable empirical risk minimization under N-way K-shot incremental sessions and the stability-plasticity dilemma; FSCIL is a subdomain of CIL but adds scarce incremental samples and sustained unreliable risk minimization.
- For FSCIC, top base-session accuracy on mini ImageNet includes SAVC (81.12%), ALICE (80.60%), BiDistFSCIL (74.65%), TEEN (73.53%), and CLOM (73.08%); on CIFAR-100, WaRP (80.31%), BiDistFSCIL (79.45%), ALICE (79.00%), SAVC (78.77%), and TEEN (74.92%); on CUB-200, SAVC (81.85%), F2M and FSIL-GAN (81.07%), DSN (80.86%), and BiDistFSCIL (79.12%).
- Lowest performance dropping (PD) on mini ImageNet: MgSvF (17.33%), SFbFSCIL (19.14%), SPPR (19.53%), ERL++ (20.94%), and DSN (21.06%); on CIFAR-100: F2M (20.04%), SFbFSCIL (20.31%), SPPR (20.65%), TEEN (22.28%), and FACT (22.50%); on CUB-200: ALICE (17.30%), DSN (17.65%), MgSvF (17.96%), TEEN (18.13%), and BiDistFSCIL (18.19%).
- For FSCIOD, top base-class methods are Sylph, iFS-RCNN, and MCH; top novel-COCO methods are iFS-RCNN, Incremental-DETR, and iMTFA; top overall COCO methods are iFS-RCNN, Sylph, and MCH; top cross-dataset VOC methods are Incremental-DETR, BPMCH, and MCH.
Methods
- Systematic literature survey of deep FSCIL in computer vision, focused on classification and object detection; covers problem definition, core challenges, general schemes, related IL/FSL problems, benchmark datasets, evaluation metrics, method taxonomy, performance comparisons, and future directions.
- Proposes taxonomy: FSCIC methods from data-based (data replay, pseudo scenarios), structure-based (dynamic structure, attention), and optimization-based (representation learning, knowledge distillation, meta-learning, other) approaches; FSCIOD methods from anchor-free (CenterNet-, FCOS-, DETR-based) and anchor-based (Mask R-CNN) frameworks.
- Performance comparison summarizes mainstream methods on mini ImageNet, CIFAR-100, and CUB-200 for classification and COCO/VOC for detection, extracting values from original papers and noting backbones and extra factors where possible.
Limitations
- The paper states FSCIL lacks publicly available datasets and that current metrics (AA, PD; mAP, mAR, mAP50) do not fully capture detailed performance throughout continuous learning.
- It states comparisons are biased by variability in backbone networks, additional data, and implementation conditions; not all methods are open-source, and performance values are extracted from original papers.
- It states the current FSCIC setting is idealistic versus real-world needs, and FSCIOD research is insufficient, with performance far from practical application and non-unified evaluation metrics. As a survey, it presents no new experiments; this is evident from the text.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
39 Valid · 16 Not valid · 27 Not checked · 134 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Despite existing surveys on FSL [27, 28, 29, 30] and IL [5, 6, 7, 11, 31, 32, 33, 34, 35], there is a clear lack of systematic and comprehensive surveys specifically on FSCIL.”
What the sources say
- Not checked — no full text of the sourcecites [27] source
- Not checked — no full text of the sourcecites [28] source
- Not checked — no full text of the sourcecites [29] source
- Supportedcites [30] source
The citations in the sentence are attached to the existence of earlier FSL and IL surveys, not to the citing authors’ conclusion about a lack of FSCIL surveys. This paper identifies itself as a survey of few-shot object detection in the Introduction and reviews that field in Section 4, so it supports its use as an example of an existing survey in a few-shot learning area. It does not establish the claimed FSCIL survey gap. (LLM, conf=0.86)
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“This approach doesn't ensure strong generalization on test data, especially with limited training samples.”
What the sources say
- Not checked — no full text of the sourcecites [27] source
- Partly supportedcites [39] sourceExperimental — accuracy not yet measured
Section 8.1 says excellent generalization is difficult with very few samples, supporting the limited-data concern. The Introduction discusses minimizing empirical risk, but our copy does not specifically say that this approach fails to ensure generalization on test data. No page numbers are available in our copy. (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“However, none of these works systematically review FSCIL.”
What the sources say
- Supportedcites [7] source
The paper identifies class-incremental learning as the subject of its survey (p. 2), and §2.3 (p. 3) narrows its evaluation to specified class-IL methods rather than few-shot class-incremental learning (FSCIL). It therefore supports the claim as applied to this cited survey, though it cannot establish the statement about every other work the citing sentence may refer to. (LLM, conf=0.84)
- Supportedcites [34] source
The Introduction describes this paper as a comprehensive survey of class-incremental learning methods, while §2.1 defines FSCIL as a distinct few-shot setting. §5 mentions a few FSCIL works as a future-direction topic rather than systematically reviewing them. The cited paper therefore supports the claim as applied to this survey; it cannot establish the scope of the other works referred to as “these works.” (LLM, conf=0.88)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“To mitigate catastrophic forgetting, some studies [44, 45, 46] use KD to maintain competent classification capabilities on previous classes while learning new ones.”
What the sources say
- Not checked — no full text of the sourcecites [44] source
- Not checked — no full text of the sourcecites [45] source
- Supportedcites [46] source
The Introduction describes UaD-CE as preserving discriminative ability on base categories while obtaining knowledge of novel categories. The Ablation Studies section explicitly says the framework uses knowledge distillation for class-incremental learning and that it helps memorize old knowledge. (LLM, conf=0.93)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Although some surveys briefly mention FSCIL [32, 35], they only provide a short introduction to the concept and a few studies without offering a thorough or systematic analysis.”
What the sources say
- Could not verifycites [32] sourceExperimental — accuracy not yet measured
The Introduction and Conclusion describe a thorough comparison of class-incremental learning algorithms, but do not characterize the paper's coverage of FSCIL. A search for “few-shot class-incremental learning” found a cited study in the References, not a passage introducing FSCIL or assessing FSCIL studies. Our copy therefore does not establish the specific claim that this survey gives FSCIL only a short introduction and a few studies. (LLM, conf=0.72)
- Supportedcites [35] source
Section 5.2, “Scarcity of Labeled Data” (page not provided), introduces FSCIL in one paragraph and briefly names representative methods. That passage supports the characterization of this survey’s FSCIL coverage as brief rather than a thorough, systematic analysis. (LLM, conf=0.88)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
56 more claims · 10 with their cited sources checked
Create a free account to see all 216 claims and to verify the 4 references still open.
- Probably not a substantial claim. 127
- It may state two separate claims as one. 74
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2.1 Problem Definition · p. 3
“| FSCIL | 2023 | Neural Networks | A Survey on Few-Shot Class-Incremental Learning [36]”
§ REFERENCES · p. 19
“- [36] S. Tian, L. Li, W. Li, H. Ran, X. Ning, and P. Tiwari, 'A survey on fewshot class-incremental learning,' Neural Networks , vol. 169, pp. 307324, 2024.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3.2.1 Evaluation Metrics for Classification · p. 7
“PD = A B - A 0”
§ 4.4.1 Performance Comparison · p. 14
“| SAVC [96] | CVPR 23 | 18 | 81.12 | 57.11 | 67.05 | 24.01”
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 1×
“as demonstrated by AlexNet [1] and ResNet [2], to the combined strategy of supervised pre-training and fine-tuning, with Transformer-based BERT [3] being a prime example.”
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×
“Therefore, this paper proposed a correction model based on Transformer [75].”
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155
Exploiting Generative AI to Scale up Intelligent Tutoring Systems
Jakubův, J., Chvalovský, K., Goertzel, Z., Kaliszyk, C., Olšák, M., Piotrowski, B., Schulz, S., Suda, M., & Urban, J. (2023) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.itp.2023.19 · Cited by 76,658
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Ren, S., He, K., Girshick, R., & Sun, J. (2017) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2016.2577031 · Cited by 73,472
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Microsoft COCO: Common Objects in Context
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460 · cited 1×
“In FSCIL, the prevalent data partitioning method by Tao et al. [13] 2) Datasets for Object Detection: COCO: The Microsoft Common Objects in Context (COCO) dataset, widely used for object detection tasks, comprises 80 object classes including people, animals, vehicles, furniture, and food [61].”
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 2×
“The core structure of this framework included the followingcomponents:First,featureextractorsobtainedthrough supervisedtrainingonbase-classdataandself-supervisedtasks onImageNet[58]orOpenImages-v6[87]usingmethodssuch aspretexttasks,contrastiveloss,orclustering.”
He, K., Gkioxari, G., Dollar, P., & Girshick, R. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.322 · Cited by 29,137
Research works printed but not linked: 2
- Kim, Warping the space: Weight space rotation for class-incremental few-shot learning
- Neal, 2012, Bayesian Learning for Neural Networks, 118
Cited by
Raza, A., Hanif, F., & Mohammed, H. A. (2025) · Scientific Reports · doi:10.1038/s41598-025-26645-2 · Cited by 56
“Transformer-based detectors (Deformable DETR, RT-DETR, PVT backbones) demonstrably improve small-object recall and occlusion robustness and can reach competitive speed–accuracy trade-offs when engineered for throughput (RT-DETR reports AP ∼ 53.1% on COCO and high FPS on T4). tasks They are particularly valuable for de…”
InfluentialYang, Y., Qu, J., Dong, W., Huang, L., & Li, Y. (2026) · IEEE Transactions on Image Processing · doi:10.1109/tip.2025.3650395 · Cited by 22
“Existing FSCIL methods can be classified into five categories: 1) feature-based methods; 2) meta-learning methods; 3) replay-based methods; 4) dynamic and prompt-based methods; and 5) regularization-based methods [30].”
Liu, Y., Jiang, T., Li, R., Yuan, L., Grzegorzek, M., Li, C., & Li, X. (2025) · Frontiers in Medicine · doi:10.3389/fmed.2025.1551894 · Cited by 16
Scale-Dependent Performance Analysis of YOLO26 and YOLOv11 for PPE Detection
Yavuz, B. Ç. (2026) · Electronics · doi:10.3390/electronics15061146 · Cited by 15
WP-FSCIL: A Well-Prepared Few-Shot Class-Incremental Learning Framework for Pill Recognition
Zhang, J., Li, C., Cristani, M., Sun, H., Grzegorzek, M., & Chen, H. (2025) · IEEE Journal of Biomedical and Health Informatics · doi:10.1109/jbhi.2025.3548691 · Cited by 10
“…which measures the absolute decline in accuracy from the first session to the last, calculated as PD = A 0 − A N , where A 0 represents the classification accuracy in the base session, and A N represents the accuracy in the final session [41]; and 3) The Average Accuracy (AA) across all sessions.”
Self-citation
Metadata
- Authors
- Zhang, J., Liu, L., Silvén, O., Pietikäinen, M., & Hu, D.
- Affiliations
- National University of Defense Technology (CN)
- CMVS, University of Oulu, Oulu, Finland
- University of Oulu (FI)
- Venue
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Published
- 2025-04
- Volume
- 47
- Issue
- 4
- Pages
- 2924-2945
- ISSN
- 0162-8828 (ISSN-L)
- 2160-9292
- 1939-3539
- OpenAlex
- W4406354592
- Semantic Scholar
- 7dbb2b74102b7bf25c2a616e68340129d6775d3a
- Cited by
- 84
- References
- 123
- Citations by year
- 2026: 58 · 2025: 24
- FWCI
- 32.63
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Natural Science Foundation of China (62036013, 62376283)
- Science and Technology Innovation Program of Hunan Province (2024QK2006)
- Key Stone (JS2023-03)
- National University of Defense Technology
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
- Keywords
- few-shot class-incremental learning, class-incremental learning, few-shot learning, object detection, classification, catastrophic forgetting, stability-plasticity dilemma, benchmark dataset, evaluation metrics, anchor-free detection, anchor-based detection
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey. https://app.cochat.ai/papers/10.1109/tpami.2025.3529038
- The paper
- Zhang, J., Liu, L., Silvén, O., Pietikäinen, M., & Hu, D. (2025). Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/tpami.2025.3529038
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.