{ "id": 304966, "title": "Grandmaster Series - How to Perform Large-Scale Image Classification", "url": "/competitions/happy-whale-and-dolphin/discussion/304966", "author_name": "", "post_date": "2022-02-03T06:54:21.771570500Z", "votes": 68, "comment_count": 6, "views": 0, "content": "

Kaggle Grandmasters share their solutions for the Google Landmark Recognition 2020 Kaggle competition.
\nIn this landmark recognition challenge, the team had to build models that recognize the correct landmark in a dataset of complicated test images. There were more than 81,000 classes in this competition.

\n

Key Takeaways:

\n
    \n
  1. ArcFace Head
    2. \n
  2. GeM pooling
    3. \n
  3. Progressive Pretraining
    4. \n
  4. Cutout Augmentation
    5. \n
\n

YouTube video: https://www.youtube.com/watch?v=VxNDH6qLZ_Q
\nEntire Playlist: Grandmaster Series

", "messages": [ { "id": "1674006", "postDate": "02/03/2022 06:54:21", "content": "

Kaggle Grandmasters share their solutions for the Google Landmark Recognition 2020 Kaggle competition.
\nIn this landmark recognition challenge, the team had to build models that recognize the correct landmark in a dataset of complicated test images. There were more than 81,000 classes in this competition.

\n

Key Takeaways:

\n
    \n
  1. ArcFace Head
    2. \n
  2. GeM pooling
    3. \n
  3. Progressive Pretraining
    4. \n
  4. Cutout Augmentation
    5. \n
\n

YouTube video: https://www.youtube.com/watch?v=VxNDH6qLZ_Q
\nEntire Playlist: Grandmaster Series

", "rawMarkdown": "Kaggle Grandmasters share their solutions for the Google Landmark Recognition 2020 Kaggle competition.\nIn this landmark recognition challenge, the team had to build models that recognize the correct landmark in a dataset of complicated test images. There were more than 81,000 classes in this competition.\n\nKey Takeaways: \n1. ArcFace Head\n2. GeM pooling\n3. Progressive Pretraining\n4. Cutout Augmentation\n\nYouTube video: [https://www.youtube.com/watch?v=VxNDH6qLZ_Q](https://www.youtube.com/watch?v=VxNDH6qLZ_Q)\nEntire Playlist: [Grandmaster Series](https://www.youtube.com/playlist?list=PL5B692fm6--uXbxtmPJz5nu3Xmc1JUm3F)", "votes": null }, { "id": "1674103", "postDate": "02/03/2022 08:30:29", "content": "

Also, you can watch this video, where the Russian team tells (with turned on subtitles) about their solution for the previous Happywhale competition:

\n

Kaggle Humpback Whale Identification: идентификация китов по изображениям — Владислав Шахрай\n

\n

\"Screenshot-44\"

", "rawMarkdown": "Also, you can watch this video, where the Russian team tells (with turned on subtitles) about their solution for the previous Happywhale competition: \n\n[Kaggle Humpback Whale Identification: идентификация китов по изображениям — Владислав Шахрай\n](https://www.youtube.com/watch?v=qwCDmpjGFA0)\n\n\"Screenshot-44\"", "votes": null }, { "id": "1674252", "postDate": "02/03/2022 10:52:21", "content": "

I would like to leave here a link to original paper that presents a GeM pooling.

\n

Meanwhile, I found another good paper about combining multiple global descriptors. The authors show that a combination of mean/global and trainable GeM pooling leads to better results. Is it kind of ensembling?

", "rawMarkdown": "I would like to leave here a link to [original paper that presents a GeM pooling](https://arxiv.org/pdf/1711.02512.pdf). \n\nMeanwhile, I found another good [paper about combining multiple global descriptors](https://arxiv.org/pdf/1903.10663.pdf). The authors show that a combination of mean/global and trainable GeM pooling leads to better results. Is it kind of ensembling?", "votes": null }, { "id": "1674404", "postDate": "02/03/2022 13:12:16", "content": "

Thanks for sharing!
\nWill have a look at them

", "rawMarkdown": "Thanks for sharing!\nWill have a look at them", "votes": null }, { "id": "1674719", "postDate": "02/03/2022 18:03:56", "content": "

There is also a blog written by them about their solution in English
\nhttps://towardsdatascience.com/a-gold-winning-solution-review-of-kaggle-humpback-whale-identification-challenge-53b0e3ba1e84

", "rawMarkdown": "There is also a blog written by them about their solution in English\nhttps://towardsdatascience.com/a-gold-winning-solution-review-of-kaggle-humpback-whale-identification-challenge-53b0e3ba1e84", "votes": null }, { "id": "1677637", "postDate": "02/05/2022 21:26:50", "content": "

Thanks for sharing this

", "rawMarkdown": "Thanks for sharing this", "votes": null }, { "id": "1678898", "postDate": "02/06/2022 20:51:28", "content": "

Thanks for the post @debarshichanda . Helpful !

", "rawMarkdown": "Thanks for the post @debarshichanda . Helpful !", "votes": null } ], "comments": [ { "id": 1674103, "author_name": "vad13irt", "author_url": "", "post_date": "02/03/2022 08:30:29", "content": "

Also, you can watch this video, where the Russian team tells (with turned on subtitles) about their solution for the previous Happywhale competition:

\n

Kaggle Humpback Whale Identification: идентификация китов по изображениям — Владислав Шахрай\n

\n

\"Screenshot-44\"

", "votes": null, "replies": [ { "id": 1674719, "author_name": "mrinath", "author_url": "", "post_date": "02/03/2022 18:03:56", "content": "

There is also a blog written by them about their solution in English
\nhttps://towardsdatascience.com/a-gold-winning-solution-review-of-kaggle-humpback-whale-identification-challenge-53b0e3ba1e84

", "votes": null, "replies": [] } ] }, { "id": 1674252, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "02/03/2022 10:52:21", "content": "

I would like to leave here a link to original paper that presents a GeM pooling.

\n

Meanwhile, I found another good paper about combining multiple global descriptors. The authors show that a combination of mean/global and trainable GeM pooling leads to better results. Is it kind of ensembling?

", "votes": null, "replies": [ { "id": 1674404, "author_name": "debarshichanda", "author_url": "", "post_date": "02/03/2022 13:12:16", "content": "

Thanks for sharing!
\nWill have a look at them

", "votes": null, "replies": [] } ] }, { "id": 1677637, "author_name": "ammarabbasi1040", "author_url": "", "post_date": "02/05/2022 21:26:50", "content": "

Thanks for sharing this

", "votes": null, "replies": [] }, { "id": 1678898, "author_name": "nebipeker", "author_url": "", "post_date": "02/06/2022 20:51:28", "content": "

Thanks for the post @debarshichanda . Helpful !

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1674006": "Kaggle Grandmasters share their solutions for the Google Landmark Recognition 2020 Kaggle competition.\nIn this landmark recognition challenge, the team had to build models that recognize the correct landmark in a dataset of complicated test images. There were more than 81,000 classes in this competition.\n\nKey Takeaways: \n1. ArcFace Head\n2. GeM pooling\n3. Progressive Pretraining\n4. Cutout Augmentation\n\nYouTube video: [https://www.youtube.com/watch?v=VxNDH6qLZ_Q](https://www.youtube.com/watch?v=VxNDH6qLZ_Q)\nEntire Playlist: [Grandmaster Series](https://www.youtube.com/playlist?list=PL5B692fm6--uXbxtmPJz5nu3Xmc1JUm3F)", "1674103": "Also, you can watch this video, where the Russian team tells (with turned on subtitles) about their solution for the previous Happywhale competition: \n\n[Kaggle Humpback Whale Identification: идентификация китов по изображениям — Владислав Шахрай\n](https://www.youtube.com/watch?v=qwCDmpjGFA0)\n\n\"Screenshot-44\"", "1674252": "I would like to leave here a link to [original paper that presents a GeM pooling](https://arxiv.org/pdf/1711.02512.pdf). \n\nMeanwhile, I found another good [paper about combining multiple global descriptors](https://arxiv.org/pdf/1903.10663.pdf). The authors show that a combination of mean/global and trainable GeM pooling leads to better results. Is it kind of ensembling?", "1674404": "Thanks for sharing!\nWill have a look at them", "1674719": "There is also a blog written by them about their solution in English\nhttps://towardsdatascience.com/a-gold-winning-solution-review-of-kaggle-humpback-whale-identification-challenge-53b0e3ba1e84", "1677637": "Thanks for sharing this", "1678898": "Thanks for the post @debarshichanda . Helpful !" }, "source": "meta" }