RFS

Rethinking Few-Shot Image Classification: A Good Embedding is All You Need?

Introduction

Name:	RFS
Embed.:	Conv64F/ResNet12/ResNet18
Type:	Fine-tuning
Venue:	ECCV'20
Codes:	rfs

• When testing the RFS-simple, you need to change test-shot to 1 or 5 for different setting with 1 checkpoint.
• Notice that in Table.2, we do not use Test-DA(which augment the test samples with 5-times) for fair.

Cite this work with:

@inproceedings{DBLP:conf/eccv/TianWKTI20,
  author    = {Yonglong Tian and
               Yue Wang and
               Dilip Krishnan and
               Joshua B. Tenenbaum and
               Phillip Isola}
  title     = {Rethinking Few-Shot Image Classification: {A} Good Embedding is All
               You Need?},
  booktitle = {Computer Vision - {ECCV} 2020 - 16th European Conference, Glasgow,
               UK, August 23-28, 2020, Proceedings, Part {XIV}},
  series    = {Lecture Notes in Computer Science},
  volume    = {12359},
  pages     = {266--282},
  year      = {2020},
  url       = {https://doi.org/10.1007/978-3-030-58568-6_16},
  doi       = {10.1007/978-3-030-58568-6_16}
}

---

Results and Models

Classification

	Embedding	📖 miniImageNet (5,1)	💻 miniImageNet (5,1)	📖miniImageNet (5,5)	💻 miniImageNet (5,5)	📝 Comments
1	ResNet12 1	62.02 ± 0.63	62.80 ± 0.52 ⬇️ 📋	79.64 ± 0.44	79.57± 0.39 ⬇️ 📋	rfs-simple-Table-1
2	Conv64F	-	47.97 ± 0.33 ⬇️ 📋	-	65.88 ± 0.30 ⬇️ 📋	Table.2
3	ResNet12	-	61.65 ± 0.35 ⬇️ 📋	-	78.88 ± 0.25 ⬇️ 📋	Table.2
4	ResNet18	-	61.65 ± 0.37 ⬇️ 📋	-	76.60 ± 0.28 ⬇️ 📋	Table.2

	Embedding	📖 tieredImageNet (5,1)	💻 tieredImageNet (5,1)	📖tieredImageNet (5,5)	💻 tieredImageNet (5,5)	📝 Comments
1	Conv64F	-	52.21 ± 0.37 ⬇️ 📋	-	71.82 ± 0.32 ⬇️ 📋	Table.2
2	ResNet12	-	70.55 ± 0.42 ⬇️ 📋	-	84.74 ± 0.29 ⬇️ 📋	Table.2
3	ResNet18	-	69.14 ± 0.42 ⬇️ 📋	-	83.21 ± 0.31 ⬇️ 📋	Table.2

Footnotes

1. ResNet12-MetaOpt with [64,160,320,640]. ↩