{ "id": 311005, "title": "Papers on Few-Shot Image Classification", "url": "/competitions/birdclef-2022/discussion/311005", "author_name": "", "post_date": "2022-03-04T10:04:54.132004Z", "votes": 22, "comment_count": 4, "views": 0, "content": "

As host says, we should classify using only few training recordings[1].

\n
\n

In this challenge, we're targeting 10 endangered Hawaiian species for which only very few training recordings are provided (due to the lack of publicly available data).

\n
\n

I think the way to properly learn with few samples are key to this competition.

\n

So I made a link list of papers on few-shot training quoted in the survey[2]. It uses four categories: 1) data augmentation, 2) embedding, 3) optimization, and 4) semantics. The number at the front is the quote number in the paper.

\n

Note that the proposed models using in these papers are not current SOTA[3]. However I think we can learn some hints from the classic techniques.

\n

1) Data Augmentation Based

\n\n

2) Embedding Based

\n\n

3) Optimization Based

\n\n

4) Semantics Based

\n\n

cf) Performance of each proposed models quoted from [2]:

\n

\"Screen-Shot-2022-03-04-at-18-28-07\"

\n

Reference

\n

[1] https://www.imageclef.org/BirdCLEF2022MainTask
\n[2] Learning from Few Samples: A Survey
\n[3] https://paperswithcode.com/sota/few-shot-image-classification-on-mini-3

", "messages": [ { "id": "1711791", "postDate": "03/04/2022 10:04:54", "content": "

As host says, we should classify using only few training recordings[1].

\n
\n

In this challenge, we're targeting 10 endangered Hawaiian species for which only very few training recordings are provided (due to the lack of publicly available data).

\n
\n

I think the way to properly learn with few samples are key to this competition.

\n

So I made a link list of papers on few-shot training quoted in the survey[2]. It uses four categories: 1) data augmentation, 2) embedding, 3) optimization, and 4) semantics. The number at the front is the quote number in the paper.

\n

Note that the proposed models using in these papers are not current SOTA[3]. However I think we can learn some hints from the classic techniques.

\n

1) Data Augmentation Based

\n\n

2) Embedding Based

\n\n

3) Optimization Based

\n\n

4) Semantics Based

\n\n

cf) Performance of each proposed models quoted from [2]:

\n

\"Screen-Shot-2022-03-04-at-18-28-07\"

\n

Reference

\n

[1] https://www.imageclef.org/BirdCLEF2022MainTask
\n[2] Learning from Few Samples: A Survey
\n[3] https://paperswithcode.com/sota/few-shot-image-classification-on-mini-3

", "rawMarkdown": "As host says, we should classify using only few training recordings[1].\n\n> In this challenge, we're targeting 10 endangered Hawaiian species for which only very few training recordings are provided (due to the lack of publicly available data).\n\nI think the way to properly learn with few samples are key to this competition.\n\nSo I made a link list of papers on few-shot training quoted in the survey[2]. It uses four categories: 1) data augmentation, 2) embedding, 3) optimization, and 4) semantics. The number at the front is the quote number in the paper.\n\nNote that the proposed models using in these papers are not current SOTA[3]. However I think we can learn some hints from the classic techniques.\n\n1) **Data Augmentation Based**\n\n- 69) [Low-shot Visual Recognition by Shrinking and Hallucinating Features](https://arxiv.org/abs/1606.02819)\n- 70) [Low-Shot Learning from Imaginary Data](https://arxiv.org/abs/1801.05401)\n- 71) [LaSO: Label-Set Operations networks for multi-label few-shot learning](https://arxiv.org/abs/1902.09811)\n- 72) [Few-Shot Learning via Saliency-guided Hallucination of Samples](https://arxiv.org/abs/1904.03472)\n- 73) [Spot and Learn: A Maximum-Entropy Patch Sampler for Few-Shot Image Classification](https://openaccess.thecvf.com/content_CVPR_2019/html/Chu_Spot_and_Learn_A_Maximum-Entropy_Patch_Sampler_for_Few-Shot_Image_CVPR_2019_paper.html)\n- 74) [Image Deformation Meta-Networks for One-Shot Learning](https://arxiv.org/abs/1905.11641)\n\n2) **Embedding Based**\n\n- 75) [Matching Networks for One Shot Learning](https://arxiv.org/abs/1606.04080)\n- 76) [Optimization as a Model for Few-Shot Learning](https://openreview.net/forum?id=rJY0-Kcll)\n- 77) [Prototypical Networks for Few-shot Learning](https://arxiv.org/abs/1703.05175)\n- 78) [Learning to Compare: Relation Network for Few-Shot Learning](https://arxiv.org/abs/1711.06025)\n- 79) [Meta-Learning for Semi-Supervised Few-Shot Classification](https://arxiv.org/abs/1803.00676)\n- 80) [TAFE-Net: Task-Aware Feature Embeddings for Low Shot Learning](https://arxiv.org/abs/1904.05967)\n- 81) [RepMet: Representative-based metric learning for classification and one-shot object detection](https://arxiv.org/abs/1806.04728)\n- 82) [Transferrable prototypical networks for unsupervised domain adaptation](https://arxiv.org/abs/1904.11227)\n- 83) [Few-shot learning with localization in realistic settings](https://arxiv.org/abs/1904.08502)\n\n3) **Optimization Based**\n\n- 84) [One-shot learning with memory-augmented neural networks](https://arxiv.org/abs/1605.06065)\n- 85) [Model-agnostic meta-learning for fast adaptation of deep networks](https://arxiv.org/abs/1703.03400)\n- 86) [Meta-SGD: Learning to Learn Quickly for Few-Shot Learning](https://arxiv.org/abs/1707.09835)\n- 87) [Deep meta-learning: Learning to learn in the concept space](https://arxiv.org/abs/1802.03596)\n- 88) [TADAM: Task dependent adaptive metric for improved few-shot learning](https://arxiv.org/abs/1805.10123)\n- 89) [Delta-encoder: an effective sample synthesis method for few-shot object recognition](https://arxiv.org/abs/1806.04734)\n- 90) [Task-Agnostic Meta-Learning for Few-shot Learning](https://arxiv.org/abs/1805.07722)\n- 91) [Meta-Transfer Learning for Few-Shot Learning](https://arxiv.org/abs/1812.02391)\n\n4) **Semantics Based**\n\n- 92) [Baby steps towards few-shot learning with multiple semantics](https://arxiv.org/abs/1906.01905)\n- 93) [Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders](https://arxiv.org/abs/1812.01784)\n- 94) [Large-Scale Few-Shot Learning: Knowledge Transfer With Class Hierarchy](https://openaccess.thecvf.com/content_CVPR_2019/papers/Li_Large-Scale_Few-Shot_Learning_Knowledge_Transfer_With_Class_Hierarchy_CVPR_2019_paper.pdf)\n\n**cf) Performance of each proposed models** quoted from [2]:\n\n\"Screen-Shot-2022-03-04-at-18-28-07\"\n\n# Reference\n\n[1] https://www.imageclef.org/BirdCLEF2022MainTask\n[2] [Learning from Few Samples: A Survey](https://arxiv.org/abs/2007.15484)\n[3] https://paperswithcode.com/sota/few-shot-image-classification-on-mini-3", "votes": null }, { "id": "1712251", "postDate": "03/04/2022 18:14:04", "content": "

Thanks for gathering these up! Anyone who decides to study the outcomes of any of these different approaches will be making an excellent contribution.

\n

A careful study of an interesting alternative can still lead to a great working note, even if it doesn't top the leaderboard. And working notes are where half the prizes are for this competition. :)

", "rawMarkdown": "Thanks for gathering these up! Anyone who decides to study the outcomes of any of these different approaches will be making an excellent contribution.\n\nA careful study of an interesting alternative can still lead to a great working note, even if it doesn't top the leaderboard. And working notes are where half the prizes are for this competition. :)", "votes": null }, { "id": "1718090", "postDate": "03/10/2022 13:52:23", "content": "

Thank you for sharing
\nUpvote

", "rawMarkdown": "Thank you for sharing \nUpvote", "votes": null }, { "id": "1718879", "postDate": "03/11/2022 08:45:01", "content": "

Below is the histogram of the sample count for the scored birds.
\nWe can see some birds have only a few sample, so we possibly require few-shot learning for these classes (unless using extra dataset).
\n\"Screen-Shot-2022-03-11-at-17-41-14\"

", "rawMarkdown": "Below is the histogram of the sample count for the scored birds.\nWe can see some birds have only a few sample, so we possibly require few-shot learning for these classes (unless using extra dataset).\n\"Screen-Shot-2022-03-11-at-17-41-14\"", "votes": null }, { "id": "1718881", "postDate": "03/11/2022 08:48:59", "content": "

I will share the result of simple EDA:
\nhttps://www.kaggle.com/tatamikenn/birdclef2022-eda-1st-step

", "rawMarkdown": "I will share the result of simple EDA:\nhttps://www.kaggle.com/tatamikenn/birdclef2022-eda-1st-step", "votes": null } ], "comments": [ { "id": 1712251, "author_name": "tomdenton", "author_url": "", "post_date": "03/04/2022 18:14:04", "content": "

Thanks for gathering these up! Anyone who decides to study the outcomes of any of these different approaches will be making an excellent contribution.

\n

A careful study of an interesting alternative can still lead to a great working note, even if it doesn't top the leaderboard. And working notes are where half the prizes are for this competition. :)

", "votes": null, "replies": [] }, { "id": 1718090, "author_name": "yasserhessein", "author_url": "", "post_date": "03/10/2022 13:52:23", "content": "

Thank you for sharing
\nUpvote

", "votes": null, "replies": [] }, { "id": 1718879, "author_name": "tatamikenn", "author_url": "", "post_date": "03/11/2022 08:45:01", "content": "

Below is the histogram of the sample count for the scored birds.
\nWe can see some birds have only a few sample, so we possibly require few-shot learning for these classes (unless using extra dataset).
\n\"Screen-Shot-2022-03-11-at-17-41-14\"

", "votes": null, "replies": [ { "id": 1718881, "author_name": "tatamikenn", "author_url": "", "post_date": "03/11/2022 08:48:59", "content": "

I will share the result of simple EDA:
\nhttps://www.kaggle.com/tatamikenn/birdclef2022-eda-1st-step

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1711791": "As host says, we should classify using only few training recordings[1].\n\n> In this challenge, we're targeting 10 endangered Hawaiian species for which only very few training recordings are provided (due to the lack of publicly available data).\n\nI think the way to properly learn with few samples are key to this competition.\n\nSo I made a link list of papers on few-shot training quoted in the survey[2]. It uses four categories: 1) data augmentation, 2) embedding, 3) optimization, and 4) semantics. The number at the front is the quote number in the paper.\n\nNote that the proposed models using in these papers are not current SOTA[3]. However I think we can learn some hints from the classic techniques.\n\n1) **Data Augmentation Based**\n\n- 69) [Low-shot Visual Recognition by Shrinking and Hallucinating Features](https://arxiv.org/abs/1606.02819)\n- 70) [Low-Shot Learning from Imaginary Data](https://arxiv.org/abs/1801.05401)\n- 71) [LaSO: Label-Set Operations networks for multi-label few-shot learning](https://arxiv.org/abs/1902.09811)\n- 72) [Few-Shot Learning via Saliency-guided Hallucination of Samples](https://arxiv.org/abs/1904.03472)\n- 73) [Spot and Learn: A Maximum-Entropy Patch Sampler for Few-Shot Image Classification](https://openaccess.thecvf.com/content_CVPR_2019/html/Chu_Spot_and_Learn_A_Maximum-Entropy_Patch_Sampler_for_Few-Shot_Image_CVPR_2019_paper.html)\n- 74) [Image Deformation Meta-Networks for One-Shot Learning](https://arxiv.org/abs/1905.11641)\n\n2) **Embedding Based**\n\n- 75) [Matching Networks for One Shot Learning](https://arxiv.org/abs/1606.04080)\n- 76) [Optimization as a Model for Few-Shot Learning](https://openreview.net/forum?id=rJY0-Kcll)\n- 77) [Prototypical Networks for Few-shot Learning](https://arxiv.org/abs/1703.05175)\n- 78) [Learning to Compare: Relation Network for Few-Shot Learning](https://arxiv.org/abs/1711.06025)\n- 79) [Meta-Learning for Semi-Supervised Few-Shot Classification](https://arxiv.org/abs/1803.00676)\n- 80) [TAFE-Net: Task-Aware Feature Embeddings for Low Shot Learning](https://arxiv.org/abs/1904.05967)\n- 81) [RepMet: Representative-based metric learning for classification and one-shot object detection](https://arxiv.org/abs/1806.04728)\n- 82) [Transferrable prototypical networks for unsupervised domain adaptation](https://arxiv.org/abs/1904.11227)\n- 83) [Few-shot learning with localization in realistic settings](https://arxiv.org/abs/1904.08502)\n\n3) **Optimization Based**\n\n- 84) [One-shot learning with memory-augmented neural networks](https://arxiv.org/abs/1605.06065)\n- 85) [Model-agnostic meta-learning for fast adaptation of deep networks](https://arxiv.org/abs/1703.03400)\n- 86) [Meta-SGD: Learning to Learn Quickly for Few-Shot Learning](https://arxiv.org/abs/1707.09835)\n- 87) [Deep meta-learning: Learning to learn in the concept space](https://arxiv.org/abs/1802.03596)\n- 88) [TADAM: Task dependent adaptive metric for improved few-shot learning](https://arxiv.org/abs/1805.10123)\n- 89) [Delta-encoder: an effective sample synthesis method for few-shot object recognition](https://arxiv.org/abs/1806.04734)\n- 90) [Task-Agnostic Meta-Learning for Few-shot Learning](https://arxiv.org/abs/1805.07722)\n- 91) [Meta-Transfer Learning for Few-Shot Learning](https://arxiv.org/abs/1812.02391)\n\n4) **Semantics Based**\n\n- 92) [Baby steps towards few-shot learning with multiple semantics](https://arxiv.org/abs/1906.01905)\n- 93) [Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders](https://arxiv.org/abs/1812.01784)\n- 94) [Large-Scale Few-Shot Learning: Knowledge Transfer With Class Hierarchy](https://openaccess.thecvf.com/content_CVPR_2019/papers/Li_Large-Scale_Few-Shot_Learning_Knowledge_Transfer_With_Class_Hierarchy_CVPR_2019_paper.pdf)\n\n**cf) Performance of each proposed models** quoted from [2]:\n\n\"Screen-Shot-2022-03-04-at-18-28-07\"\n\n# Reference\n\n[1] https://www.imageclef.org/BirdCLEF2022MainTask\n[2] [Learning from Few Samples: A Survey](https://arxiv.org/abs/2007.15484)\n[3] https://paperswithcode.com/sota/few-shot-image-classification-on-mini-3", "1712251": "Thanks for gathering these up! Anyone who decides to study the outcomes of any of these different approaches will be making an excellent contribution.\n\nA careful study of an interesting alternative can still lead to a great working note, even if it doesn't top the leaderboard. And working notes are where half the prizes are for this competition. :)", "1718090": "Thank you for sharing \nUpvote", "1718879": "Below is the histogram of the sample count for the scored birds.\nWe can see some birds have only a few sample, so we possibly require few-shot learning for these classes (unless using extra dataset).\n\"Screen-Shot-2022-03-11-at-17-41-14\"", "1718881": "I will share the result of simple EDA:\nhttps://www.kaggle.com/tatamikenn/birdclef2022-eda-1st-step" }, "source": "meta" }