PaperlayerView exact revision on arXiv

Source-linked AI summary

Few-Shot Object Detection: A Comprehensive Survey

Mona Köhler, Markus Eisenbach, Horst-Michael Gross

arXiv:2112.11699v2cs.CVcs.LG

TL;DR

Deep-learning object detection typically needs extensive annotations, while novel categories may have few or no obtainable examples. This survey categorizes FSOD approaches, reviews datasets and benchmark protocols, and analyzes reported results. Its benchmark analysis identifies trends among strong approaches, but leaves open which of two leading approaches will prevail.

  • Problem

    Deep-learning detectors require extensive annotated data, but detecting novel categories is difficult when only few target-domain instances are available or annotations are costly or impossible to obtain.

    Source: §I · p. 1

  • Method

    The survey categorizes FSOD by training scheme and architecture, describes approaches and their concepts, and reviews datasets, evaluation protocols, and benchmark results.

    Source: §I · p. 1, §XI · p. 20

  • Results

    Analysis of PASCAL VOC and Microsoft COCO benchmarks identifies trends in the best-performing dual-branch meta-learning and transfer-learning approaches, but which will prevail remains open.

    Source: §XI · p. 20

  • Takeaways & Limitations

    The survey emphasizes evaluation challenges and identifies promising research directions to guide future work on FSOD.

    Source: §I · p. 1

Abstract

from arXiv · show

Humans are able to learn to recognize new objects even from a few examples. In contrast, training deep-learning-based object detectors requires huge amounts of annotated data. To avoid the need to acquire and annotate these huge amounts of data, few-shot object detection aims to learn from few object instances of new categories in the target domain. In this survey, we provide an overview of the state of the art in few-shot object detection. We categorize approaches according to their training scheme and architectural layout. For each type of approaches, we describe the general realization as well as concepts to improve the performance on novel categories. Whenever appropriate, we give short takeaways regarding these concepts in order to highlight the best ideas. Eventually, we introduce commonly used datasets and their evaluation protocols and analyze reported benchmark results. As a result, we emphasize common challenges in evaluation and identify the most promising current trends in this emerging field of few-shot object detection.

I. INTRODUCTION

Deep-learning detectors need extensive annotations, making novel-object detection costly or infeasible in some domains; FSOD uses abundant base-category data to detect novel categories from few annotated instances.

II. PROBLEM DEFINITION

  • FSOD separates abundant base-category data from novel categories with exactly K annotated object instances each; training only on novel data risks overfitting, while combined-data training can bias detection toward base categories.A single image may contain multiple annotated instances, so K is not necessarily the number of images.Source: §II · pp. 1–2

A. Related Concepts for Learning with Limited Data

  • Few-shot classification ideas,,, can transfer to detection, while semi-supervised learning uses additional unlabeled data,, and incremental learning seeks to retain old-category performance,,.Few-shot detection differs from classification by also requiring localization.Source: A. Related Concepts for Learning with… · p. 2

B. Object Detection

  • Cross-domain detection adapts to a new domain while keeping categories fixed,,; zero-shot detection uses K=0, and weakly supervised detection uses image-level labels,,, [26].Weakly supervised methods still require many images, which may be hard to obtain for rare objects.Source: B. Object Detection · p. 2

C. Learning Techniques for Few-Shot Object Detection

  • Transfer learning reuses pretrained weights but may involve more target instances than few-shot learning, while metric and meta learning seek representations or learning procedures that generalize to novel categories.Metric-based few-shot detection must add localization to feature-based matching; meta learning trains models to learn category knowledge from few examples.Source: C. Learning Techniques for Few-Shot… · pp. 2–3

D. Related Surveys

The survey organizes FSOD by training scheme and architecture, then describes meta-learning realizations and variants alongside transfer-learning approaches. It also compares benchmark results while distinguishing evaluation protocols.

D. Related Surveys

  • Compared with prior surveys, this work covers nearly twice as many FSOD papers as [32] and groups benchmark results by comparable evaluation protocols.It also uses visual taxonomies to show which concepts appear similar or complementary.Source: D. Related Surveys · p. 3

IV. CATEGORIZATION OF FEW-SHOT OBJECT DETECTION APPROACHES

  • The survey categorizes FSOD by training scheme and architecture, dividing meta learning into single- and dual-branch approaches rather than treating metric learning as a separate category.The authors argue this better reflects current trends.Source: §IV · p. 3
  • The survey reports that most covered approaches use transfer learning or dual-branch meta learning, and that most were published within the preceding three years.It summarizes approaches, takeaways, and best-performing methods, then compares training schemes and benchmark results.Source: §IV · p. 3

A. Training Scheme

  • Meta learning trains on base data, uses episodes that mimic N-way-K-shot tasks, and then either finetunes on the final task or meta-tests with annotated novel examples.For detecting both base and novel categories, finetuning uses K examples per category from each; novel-only settings use novel examples.Source: A. Training Scheme · pp. 3–4

B. General Realization

  • Dual-branch methods process a query image and a support set separately, then aggregate support features with query features to guide detection toward the supported category.Support examples may be cropped, supplied with a mask, or represented by features extracted from their designated object regions.Source: B. General Realization · p. 4
  • In the common Faster R-CNN realization, pooled support vectors specialize query RoIs by category before a shared RoI head performs box regression and binary classification.The shared head must generalize across categories, allowing meta-testing on novel categories without finetuning; during inference, support features can be computed once for all categories.Source: B. General Realization · pp. 4–5

C. Variants for Aggregation

Dual-branch approaches vary how query and support features are combined, including aggregation before the region proposal network and different feature-comparison operations.

1) Aggregation before the Region Proposal Network:

  • PNSD replaces average pooling with second-order pooling and power normalization to capture feature co-occurrences and reduce variability from appearance differences.KFSOD uses kernelized covariance matrices and RKHS kernels to capture nonlinear patterns while retaining regional statistics and tolerating spatial shifts.Source: 1) Aggregation before the Region… · p. 5
  • Before the RPN, AttentionRPN aggregates query and support features to produce category-related proposals, improving recall and reducing missed detections.Support features are average-pooled, then combined with query features by depth-wise cross-correlation.Source: 1) Aggregation before the Region… · p. 5

2) Aggregation Operation:

  • Aggregation operations include channel-wise multiplication, convolution or correlation, cosine similarity, and position-wise ℓ1 distance.FsDetView combines multiplication, subtraction, and query features; its component functions are small convolutional networks.Source: 2) Aggregation Operation · p. 5
  • SQMG uses support features to generate dynamic-convolution weights that are applied to query features, while simple channel-wise multiplication cannot fully exploit both feature sets.Source: 2) Aggregation Operation · p. 5

3) Keep Spatial Information for Aggregation:

Aggregation methods preserve support-image spatial information, use attention to relate query and support features, interact across feature levels, and account for differences among support images.

3) Keep Spatial Information for Aggregation:

  • Spatial information can be lost through global average pooling, which also mixes irrelevant box features; OSWF instead compares spatially pooled support and query features with 1×1 convolutions.Convolution can be unsuitable when query and support objects are spatially misaligned.Source: 3) Keep Spatial Information for… · p. 6

4) Attention-based Aggregation:

  • DAnA uses dual-awareness attention to highlight category-relevant support features before modeling spatial correlations, while Meta Faster R-CNN uses attention to align support and RoI features and highlight foreground regions.Source: 4) Attention-based Aggregation · p. 6
  • IFC combines self-attention with a learnable soft-threshold operator to reduce redundant information and enhance feature sensitivity and stability for base and novel categories.ARRM integrates global support context and uses attention-based affinity reasoning to reduce misclassifications.Source: 4) Attention-based Aggregation · p. 6
  • CoAE uses mutual non-local operations for co-attention, while AIT uses multi-head co-attention before the RPN and reports improved accuracy over CoAE’s non-local block.AIT’s resulting features encode related query and support characteristics to help the RPN predict relevant RoIs.Source: 4) Attention-based Aggregation · p. 6

5) Multi-Level Aggregation:

  • FCT interacts between query and support features at three backbone stages and one detection-head stage, then uses pairwise matching to produce detections.The approach is motivated by aligning features through multi-level interaction.Source: 5) Multi-Level Aggregation · p. 7

6) Aggregation of Several Support Images:

  • GenDet learns weights from similarity between each support image’s single-shot and mean detector, while CAReD weights images using softmax-normalized correlations with the other same-category support features.Source: 6) Aggregation of Several Support Images · p. 7
  • Because support images can differ in information, their individual relevance should be incorporated rather than treating them as equally informative.SQMG weights support features by query similarity; APSP refines individual support vectors and attends over them with query features.Source: 6) Aggregation of Several Support Images · p. 7

D. Incorporate Relations between Categories

Category relations can provide information for classifying novel categories, especially when they resemble base categories; surveyed approaches incorporate those relations through linguistic or visual features.

D. Incorporate Relations between Categories

  • Relations between categories can aid classification when novel categories resemble base categories, unlike approaches that treat each category independently.Source: D. Incorporate Relations between… · pp. 7–8
  • FSOD-KT transfers semantic correlations through a graph convolutional network over category-name similarities, but requires predefined distinct names and may be difficult to apply to medical imaging.Source: D. Incorporate Relations between… · p. 7
  • QA-FewDet builds a heterogeneous graph to model multi-category relations and adapt support and query features, while Meta-DETR attends to multiple support categories to capture inter-class correlations.Meta-DETR’s correlation aggregation is described as helping reduce misclassification and enhance generalization to novel categories.Source: D. Incorporate Relations between… · pp. 7–8

E. Increase Discriminative Power

The survey describes ways to improve category discrimination and representation, use proposal-free detectors, and retain performance on base categories during novel-category learning.

E. Increase Discriminative Power

  • GenDet and Meta-DETR constrain cosine similarities between support vectors toward an identity matrix, while other approaches use cosine or Pearson distances, margins, or adversarial feature disturbance.GenDet also uses a reference detector to distinguish all base categories across episodes, and CoAE and AIT use a margin-based ranking loss for RPN proposals.Source: E. Increase Discriminative Power · pp. 8–9
  • To distinguish categories and avoid false positives when the target category is absent, approaches use discriminative support features, similarity or ranking losses, contrastive training, and hard-negative-focused focal loss.The survey highlights metric-learning and contrastive-training ideas, including two-way contrastive training with negative support images from categories absent in the query.Source: E. Increase Discriminative Power · pp. 8–9

F. Improve representation capability

G. Proposal-free Detectors

  • Most approaches use Faster R-CNN, whereas proposal-free detectors allow support and query features to be aggregated after feature extraction; the survey reports that transformer-based Meta-DETR already surpasses other approaches.Source: G. Proposal-free Detectors · p. 9

H. Keep the Performance on Base Categories

  • Meta Faster R-CNN protects base-category performance by keeping its base branch fixed during meta-finetuning, while incremental approaches keep weights for learned categories fixed.Sylph uses independent sigmoid-based category classifiers and a support-branch hypernetwork to generate novel-category classifier weights, avoiding meta-finetuning.Source: H. Keep the Performance on Base… · p. 9

I. Increase the Variance of Novel Categories

The survey groups approaches around support-example augmentation, context use, shared box regression, and metric learning, then highlights leading dual-branch methods and their complex episodic-training requirement.

I. Increase the Variance of Novel Categories

  • TIP uses consistency loss to keep support features from original and transformed images close, while transformed query proposals are matched to RoIs from the original image.This encourages representative support vectors and transformation-independent query proposals.Source: §I · p. 9

J. Incorporate Context Information

  • DCNet pools RoIs at three resolutions and fuses branches with attention, while QA-FewDet uses graph edges to add local and global context to noisy proposals.Larger pooling resolutions capture local detail and smaller ones holistic information.Source: J. Incorporate Context Information · p. 9

K. Category-agnostic Bounding Box Regression

Summary of Best Performing Dual-Branch Meta Learning Approaches

  • Meta-DETR uses category-aware attention and cosine-similarity loss; FCT aggregates support and query features across levels; IFC adds interactive attention and orthogonal cosine loss; SQMG uses mutual guidance and classification-focused training.SQMG does not require finetuning and shares box regression across categories.Source: Summary of Best Performing Dual-Branch… · p. 10

Conclusion on Dual-Branch Meta Learning Approaches

VI. SINGLE-BRANCH META LEARNING

  • Single-branch meta-learning resembles generic detectors, but its approaches vary and all use episodic training.Source: §VI · p. 10

A. Metric Learning

  • RepMet, NP-RepMet, and PNPDet use metric learning to form discriminative embeddings, with negative proposals or category prototypes supporting novel-category recognition.The survey's takeaway is that metric learning helps distinguish categories.Source: A. Metric Learning · p. 10

B. Reduce Learnable Parameters

The survey describes parameter-efficient single-branch meta-learning strategies and transfer learning, which uses a simpler two-phase setup, then reviews RPN changes for better novel-category proposals.

B. Reduce Learnable Parameters

  • Freezing category-agnostic detector weights or adding a small novel-category classifier reduces learnable parameters when finetuning on scarce novel-category data.MetaDet predicts category-specific weights episodically; MetaRetinaNet learns coefficient vectors; PNPDet adds a classifier while freezing the network.Source: B. Reduce Learnable Parameters · p. 11

C. Learnable Learning Rate

  • Meta-SSD learns individual parameter learning rates from each task's distribution, enabling adaptation to novel categories with one parameter update.Source: C. Learnable Learning Rate · p. 11

D. Balanced Loss Function

  • MetaRetinaNet's balancing loss constrains performance across episodes because episode-to-episode imbalance hinders stable adaptation to novel categories.Source: D. Balanced Loss Function · p. 11

Conclusion on Single-Branch Meta Learning Approaches

VII. TRANSFER LEARNING

  • Transfer learning trains on base categories, freezes all weights except the RoI head, then finetunes on balanced base and novel data; cosine classification compensates for feature-norm differences.The survey reports that this simple approach outperformed earlier, more complex meta-learning approaches.Source: §VII · p. 11

A. Modifications of the Region Proposal Network

  • CoRPN replaces one foreground classifier with multiple binary classifiers, while other approaches adapt RPN weights during finetuning; FSCE and CIR also increase post-NMS proposals.FSCE samples half as many proposals in the RoI head because the discarded half were observed to contain only backgrounds.Source: A. Modifications of the Region… · p. 12

B. Modifications of the Feature Pyramid Network

The survey covers adapting the feature pyramid, increasing novel-category example variance, transferring knowledge from base categories, retaining base-category performance, and modifying training objectives.

B. Modifications of the Feature Pyramid Network

  • Finetuning the FPN improves performance over freezing it, while context modules and augmentation address scale variation and small-object detection.MPSR augments object crops at different scales; CIR enlarges the FPN receptive field.Source: B. Modifications of the Feature… · p. 12

C. Increase the Variance of Novel Categories

  • Augmentation and pseudo-labeling increase novel-category example variance and improve detection, especially when K is very low.LVC verifies and corrects pseudo-labeled boxes, but real-world applications may require additional unlabeled data.Source: C. Increase the Variance of Novel… · pp. 12–13

D. Transfer Knowledge Between Base and Novel Categories

  • Approaches transfer information from semantically related base categories through weight initialization, visual-semantic representations, semantic associations, or knowledge graphs.FADI associates each novel category with one most-similar base category for initialization, while noting this can cause base/novel confusion.Source: D. Transfer Knowledge Between Base and… · p. 13

E. Keep the Performance on Base Categories

  • To limit catastrophic forgetting, methods separate base and novel components, retain base-category representatives, or adjust updates using the angle between base and novel gradients.CFA's episodic memory stores K base-category shots for computing base gradients during finetuning.Source: E. Keep the Performance on Base… · p. 13

F. Modify the Training Objective

  • Training objectives are modified to focus on foreground regions, use source-domain knowledge, address unlabeled instances, and improve inter-class separability.Source: F. Modify the Training Objective · p. 14

1) Additional Loss Terms:

The survey reviews losses, auxiliary branches, gradient-flow and training-scheme changes, attention, and detector architectures as ways to adapt transfer-learning approaches for few-shot detection.

1) Additional Loss Terms:

  • Additional loss terms address background suppression, knowledge transfer, unlabeled instances, class similarity, inter-class margins, and uncertainty in pseudo-labels.Source: 1) Additional Loss Terms · p. 14

2) Loss for Auxiliary Branches:

  • Auxiliary branches support consistency, foreground segmentation, adversarial training, or contrastive and metric-learning objectives; FSSP also combines original and augmented-branch decisions.Source: 2) Loss for Auxiliary Branches · p. 14

3) Modified Gradient Flow:

  • DeFRCN stops RPN gradients to the backbone and scales RoI-head gradients by 0.75 in base training and 0.01 in base-plus-novel training, significantly boosting performance.The approach targets the conflicting RPN and RoI-head objectives.Source: 3) Modified Gradient Flow · p. 14

4) Modified Training Scheme:

  • Staged single-shot then K-shot finetuning and TD-Sampler's difficulty-guided batches are proposed to alter training and sample use.TD-Sampler uses all base samples and more novel samples per batch while limiting changes to detector activation patterns.Source: 4) Modified Training Scheme · p. 14

G. Use Attention

  • Attention methods enhance features or incorporate context, including prototype-to-RPN attention, relational attention, self-attention, and top-down plus bottom-up attention.Source: G. Use Attention · pp. 14–15

1) Architectures based on Faster R-CNN:

2) Incorporating One-Stage Detectors:

  • FSSP makes YOLOv3 competitive with vanilla Faster R-CNN using self-attention, a replicated auxiliary detector branch, augmentation, and an additional loss; it performs better than TFA in the extremely low-shot setting.The survey takeaway is that a single-stage detector can benefit from an auxiliary augmentation branch.Source: 2) Incorporating One-Stage Detectors · p. 15

Summary of Best Performing Transfer Learning Approaches

The survey highlights transfer-learning techniques for few-shot detection, compares them with meta-learning, and reviews PASCAL VOC protocols and results. It identifies evaluation variability and dataset difficulty as important qualifications.

Summary of Best Performing Transfer Learning Approaches

  • DeFRCN controls gradient flow between the RPN, ROI head, and backbone, then uses prototypical calibration to refine scores and reduce high-scored false positives.The survey reports that training all detector components in both phases significantly boosts performance, especially during finetuning.Source: Summary of Best Performing Transfer… · p. 15
  • CFA stores K base-category examples and adaptively combines base and novel gradients when their angle is obtuse, addressing forgetting during finetuning.The update rule also benefits novel-category performance.Source: Summary of Best Performing Transfer… · pp. 15–16
  • FADI aligns novel-category features with a semantically similar base category, then disentangles their classification branches and applies a margin loss to improve separation.The initial alignment reduces novel-category intra-class variation but can confuse base and novel categories.Source: Summary of Best Performing Transfer… · p. 16
  • LVC verifies pseudo-labeled novel samples using vision-transformer features, corrects their boxes, and uses the extended data for another finetuning phase.The approach is intended for settings with additional unlabeled data containing novel categories.Source: Summary of Best Performing Transfer… · p. 16

Conclusion on Transfer Learning Approaches

  • Transfer learning avoids complex episodic training and can reach state-of-the-art performance by enabling more detector components to be finetuned.The survey cites modified objectives and knowledge transfer between base and novel categories as examples.Source: Conclusion on Transfer Learning… · p. 16

VIII. COMPARISON BETWEEN META LEARNING AND TRANSFER LEARNING

  • The comparison excludes less-explored single-branch meta-learning; dual-branch meta-learning and transfer learning both remain promising and could benefit from exchanging ideas.The survey compares the two schemes across several aspects in Table I.Source: §VIII · p. 16, Table I · p. 16

A. PASCAL VOC Dataset

  • PASCAL VOC commonly uses three splits with 15 base and five novel categories, K values of 1, 2, 3, 5, and 10, and AP50; variable sampled instances complicate comparisons, while VOC can permit localization without informative support.The revised protocol averages results across 30 random runs and reports base-category performance; generalized FSOD averages novel- and base-category performance.Source: A. PASCAL VOC Dataset · pp. 16–18, §IX · p. 16
  • On PASCAL VOC, dual-branch meta-learning and transfer-learning approaches achieve similar results, while averaging runs yields more reliable comparisons.The benchmark table distinguishes single-run from multi-run results.Source: Table II · p. 17, A. PASCAL VOC Dataset · p. 18

B. Microsoft COCO Dataset

The survey describes Microsoft COCO protocols and evaluation pitfalls, then identifies current technical trends and open challenges. It notes that common benchmarks do not represent realistic rare novel categories and that realistic settings may involve domain shift.

B. Microsoft COCO Dataset

  • COCO few-shot detection commonly treats the 20 VOC categories as novel and the other 60 as base, using 10 or 30 shots, with 1–3 shots in some extreme low-shot studies.Evaluation uses COCO AP metrics, including AP50∶95 as the primary metric, and published 10-shot results are listed in Table III.Source: B. Microsoft COCO Dataset · p. 18, Table III · p. 19

C. Deviating Evaluation Protocols

  • Some one-shot and one-way protocols assume the query contains the support category, removing classification; the survey therefore recommends N-way evaluation for comparability.PNSD and KFSOD also use a COCO-pretrained ResNet, so their novel categories are not truly novel.Source: C. Deviating Evaluation Protocols · p. 18

1) High Variance for Different Samples:

  • Different novel-category samples can produce high performance variance and hinder comparisons, so the survey recommends averaging results over multiple random runs.The cited recommendation is attributed to Wang et al..Source: 1) High Variance for Different Samples · p. 18

2) ImageNet Pretraining and Choice of Novel Categories:

  • ImageNet pretraining may expose detectors to novel categories, while omitting it worsens base-category performance; proposed remedies exclude corresponding pretraining classes or use rare categories absent from ImageNet.Removing all COCO-related classes may reduce pretraining representational power, and class exclusions can require separate pretraining for each novel-category set; LVIS rare categories also have validation-set coverage concerns.Source: 2) ImageNet Pretraining and Choice of… · pp. 18–19

X. CURRENT TRENDS

  • Current methods use attention to aggregate dual-branch features, aggregate before the RPN or use proposal-free transformers to address missing proposals, and guide gradients to train more detector components.Metric-learning losses are also a current trend for improving category separation.Source: §X · p. 20
  • FSOD concepts are also being extended to weakly supervised and self-supervised settings and to few-shot instance segmentation.Source: §X · p. 20
  • The survey calls for evaluating on realistic rare-category datasets such as LVIS or FSOD, exploring cross-domain detection, and exchanging concepts across meta-learning and transfer-learning approaches.A domain shift is likely when FSOD is applied to realistic settings with truly rare categories; which of the two training schemes will prevail remains open.Source: §X · p. 20, §XI · p. 20

APPENDIX

The appendix reviews generic object-detection architectures and notation, then outlines the survey’s scope. It describes a field in which most covered approaches use transfer learning or dual-branch meta-learning.

A. Generic Object Detection

  • One-stage detectors classify and regress boxes directly from backbone feature maps, whereas two-stage detectors first generate object proposals and then classify and refine them.Many few-shot detectors build on Faster R-CNN, often with a ResNet backbone.Source: A. Generic Object Detection · p. 23
  • In Faster R-CNN, a backbone feeds an RPN that proposes regions; RoI pooling standardizes their feature resolution before the box head classifies and refines them.The RPN classifies anchors as foreground or background and regresses anchor deltas.Source: A. Generic Object Detection · p. 23, Figure 11 · p. 23
  • An FPN supplies feature maps at multiple resolutions, with the RPN applied at each scale and RoI pooling matched to proposal size.The survey presents the FPN architecture in Fig. 12.Source: A. Generic Object Detection · p. 23, Figure 12 · p. 23

B. Mathematical notation

  • The notation reference maps symbols to backbone, RPN, RoI, query, support, category, and dataset concepts used throughout the survey.It identifies K as the number of shots, N as the number of novel categories, and Minit, Mbase, and Mfinal as successive detector models.Source: B. Mathematical notation · p. 23

C. Scope of this survey

D. Further benchmark results

The section reviews benchmark settings and dataset options for few-shot object detection, then distinguishes this survey by its focused coverage, organization, and benchmark guidance.

1) FSOD dataset:

  • The FSOD dataset has 1,000 categories, with 800 base and 200 novel categories, and its proposed 2-way-5-shot protocol does not finetune on novel data.Its novel categories are selected to be distinct from base categories and exclude COCO categories; only four approaches had been benchmarked there so far.Source: 1) FSOD dataset · p. 24, Table IV · p. 24

2) PASCAL VOC with pretraining on reduced ImageNet:

3) Microsoft COCO:

  • The section reports Microsoft COCO benchmark results for K = 1 and K = 30, with the 30-shot table sorted by novel AP50∶95.The tables mark results with deviating evaluation protocols as unsuitable for fair comparison and identify approach families with symbols.Source: 3) Microsoft COCO · p. 25, Table V · p. 25, Table VIII · p. 27

4) Other datasets:

  • RepMet’s ImageNet-LOC benchmark trains on 100 mostly animal categories and evaluates on 214 unseen animal categories in a 5-way setting.Source: 4) Other datasets · p. 25
  • LVIS rare categories, with at most 10 training examples, can serve as novel categories, but their rarity in validation makes evaluation prone to high variance.The split has been used by TFA and LST so far.Source: 4) Other datasets · p. 25

E. Relation to other surveys

  • Unlike broader surveys, this survey focuses on FSOD to provide more detailed insight into its techniques and trends.The surveys cited as broader also cover self-supervised, weakly supervised, or zero-shot learning.Source: E. Relation to other surveys · p. 25
  • It organizes categories by general realization and then discusses differing techniques, a structure the authors say makes trends easier to derive.Source: E. Relation to other surveys · p. 26
  • The survey covers more FSOD papers than the related surveys, including approaches from the current year, and reports nearly twice as many as [32].The authors note that and [31] omit some currently best-performing approaches on common benchmarks.Source: E. Relation to other surveys · p. 26
Loading 2112.11699v2…