CoChat
Sign Up
Cited by 56Domain Adaptation and Few-Shot LearningAdvanced Neural Network ApplicationsAdvanced Image and Video Retrieval Techniques

Sylph: A Hypernetwork Framework for Incremental Few-shot Object Detection

Open access
Yin, Li, Perez-Rua, Juan M, Liang, Kevin J·2022·2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
DOI 10.1109/cvpr52688.2022.00883
Open this paper in CoChat — free accountOpen PDF

Abstract

We study the challenging incremental few-shot object de-tection (iFSD) setting. Recently, hypernetwork-based approaches have been studied in the context of continuous and finetune-free iFSD with limited success. We take a closer look at important design choices of such methods, leading to several key improvements and resulting in a more accurate and flexible framework, which we call Sylph. In particular, we demonstrate the effectiveness of decou-pling object classification from localization by leveraging a base detector that is pretrained for class-agnostic local-ization on large-scale dataset. Contrary to what previous results have suggested, we show that with a carefully de-signed class-conditional hypernetwork, finetune-free iFSD can be highly effective, especially when a large number of base categories with abundant data are available for meta-training, almost approaching alternatives that undergo test-time-training. This result is even more significant considering its many practical advantages: (1) incrementally learning new classes in sequence without additional training, (2) detecting both novel and seen classes in a single pass, and (3) no forgetting of previously seen classes. We benchmark our model on both COCO and LVIS, reporting as high as 17% AP on the long-tail rare classes on LVIS, indicating the promise of hypernetwork-based iFSD.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1109/cvpr52688.2022.00883

Cites

11 of 11 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Anchor-free frameworks can effectively handle incremental classes without defining anchor boxes.

    “The reason is that anchor-free frameworks can effectively handle incremental classes without defining anchor boxes.”

    From Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey (2025) · cites this paper as [99]

    Section 3.1.1 describes FCOS as an anchor-free detector used to adapt to novel classes, supporting the claim for this framework. But Section 2 attributes incremental class handling to class-agnostic localization and separate per-class classifiers; the paper does not establish that omitting anchor boxes is itself the reason anchor-free frameworks handle incremental classes effectively. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Cites

Flagged first, then by citations

Research works printed but not linked: 61

  1. zhang, 2021, Meta-detr Image-level few-shot object detection with inter-class correlation exploitation
  2. yuewei, 2020, Object detection as a positive-unlabeled problem, In BMVC
  3. 10.1109/ICCV.2019.00967 · DOI
  4. liu, 2016, Ssd: Single shot multibox detector, ECCV, 21
  5. 10.1109/CVPR52688.2022.00888 · DOI
  6. 10.1109/CVPR46437.2021.00728 · DOI
  7. sachin, 2022, Extending one-stage detection with open-world proposals
  8. lin, 2014, Microsoft COCO: Common Objects in Context, ECCV, 740
  9. 10.1109/ICCV.2017.324 · DOI · Paper page
  10. flood, 2018, Learning to compare: Re-lation network for few-shot learning, CVPR, 1199
  11. thrun, 2012, Learning to learn, Springer Science & Business Media
  12. vanschoren, 2018, Meta-Learning A Survey
  13. 10.1109/ICCV.2019.00851 · DOI
  14. vinyals, 2016, Matching networks for one shot learning, NeurIPS, 29, 3630
  15. joseph, 2021, Towards open world object de-tection, CVPR, 5830
  16. xin, 2020, Frustratingly simple few-shot object detection
  17. wang, 2019, Tafe-net: Task-aware feature embed-dings for low shot learning, CVPR, 1831
  18. kim, 2021, Learning open-world ob-ject proposals without learning to classify
  19. 10.1007/978-3-030-01261-8_1 · DOI · Paper page
  20. wu, 2019, Detectron2
  21. xiao, 2020, Few-shot object detection and viewpoint estimation for objects in the wild, ECCV, 192
  22. bansal, 2018, Zero-shot object detection, ECCV, 384
  23. manoj, 2020, Rodeo: Replay for online object detection, BMVC
  24. 10.1109/CVPR52688.2022.01842 · DOI · Paper page
  25. gu, 2021, Zero-shot detection via vision and language knowledge distillation
  26. he, 2017, Mask r-cnn, ICCV, 2961
  27. loffe, 2015, Batch normalization: Accelerating deep network training by reducing internal covariate shift, ICML, 448
  28. 10.1109/CVPR.2017.351 · DOI · Paper page
  29. rangrej, 2021, Revisiting linear decision boundaries for few-shot learning with transformer hypernetworks
  30. 10.1037/0033-295X.97.2.285 · DOI
  31. sun, 2020, Test-time training with self-supervision for generalization under distribution shifts, ICML
  32. snell, 2017, Prototypi-cal networks for few-shot learning, NeurIPS, 30, 4077
  33. simonyan, 2014, Very Deep Convolutional Networks for Large-scale Image Recognition
  34. salimans, 2016, Weight normalization: A simple reparameterization to accelerate training of deep neural networks, NeurIPS, 29, 901
  35. 10.1007/s11263-015-0816-y · DOI · Paper page
  36. ravi, 2017, Optimization as a model for few-shot learning, ICLRE
  37. 10.1109/CVPR42600.2020.00407 · DOI
  38. 10.1109/CVPR46437.2021.00450 · DOI
  39. nikhil, 2021, Continual learning using a bayesian non-parametric dictionary of weight factors, International Conference on Artificial Intelligence and Statistics, 100
  40. farquhar, 2018, Towards ro-bust evaluations of continual learning
  41. finn, 2017, Model- agnostic meta-learning for fast adaptation of deep networks, ICML, 1126
  42. garnelo, 2018, Conditional neu-ral processes, ICML, 1704
  43. 10.1109/CVPR46437.2021.00294 · DOI
  44. 10.1109/CVPR.2018.00459 · DOI · Paper page
  45. 10.1109/ICCV.2015.169 · DOI · Paper page
  46. 10.1109/CVPR.2014.81 · DOI · Paper page
  47. goodfellow, 2013, An empirical investigation of catas-trophic forgetting in gradient-based neural networks
  48. chen, 2021, Dual-awareness attention for few-shot ob-ject detection
  49. luca, 2016, Learning feed-forward one-shot learners, NeurIPS, 523
  50. dai, 2016, R-fcn: Object detection via region-based fully convolutional networks, NeurIPS, 379
  51. 10.1109/CVPRW50498.2020.00359 · DOI · Paper page
  52. 10.1109/ICCV.2019.00667 · DOI · Paper page
  53. ramanathan, 2021, Pre-det: Large-scale weakly supervised pre-training for detection, ICCV, 2865
  54. 10.1007/s11263-009-0275-4 · DOI · Paper page
  55. pinker, 2003, How the mind works, Penguin UK
  56. 10.1109/CVPR.2018.00755 · DOI
  57. 10.1109/ICCV48922.2021.00856 · DOI
  58. 10.1016/j.neunet.2019.01.012 · DOI · Paper page
  59. nichol, 2018, On first-order meta-learning algorithms
  60. perez-rua, 2020, Incremental few-shot ob-ject detection, CVPR, 13846
  61. pascanu, 2013, On the difficulty of training recurrent neural networks, ICML, 1310

Cited by

Sorted by citations

Authors
Yin, L., Perez-Rua, J. M., & Liang, K. J.
Affiliations
  • Meta AI
  • Meta (United States) (US)
ORCID
Liang, Kevin J
Venue
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Publisher
IEEE
Published
2022-06
Pages
9025-9035
ISSN
1063-6919 (ISSN-L)
DOI
10.1109/cvpr52688.2022.00883
OpenAlex
W4312530527
Cited by
56
References
11
Citations by year
2026: 4 · 2025: 11 · 2024: 15 · 2023: 24 · 2022: 2
FWCI
7.19
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
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 › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
Keywords
hypernetworks, meta-training

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.

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