{ "id": 314096, "title": "How species in the train data is selected?", "url": "/competitions/birdclef-2022/discussion/314096", "author_name": "", "post_date": "2022-03-21T01:39:39.843246600Z", "votes": 8, "comment_count": 4, "views": 0, "content": "

@stefankahl @tomdenton @amandanavine

\n

I have a question to the host.

\n

I downloaded the list of bird species shown in Hawaii[1], and compared it to species in train data.
\nFrom the data, I find out most of the species are in the list, but some are not in the list(Fig. 1).
\nBesides, the species in the train dataset doesn't cover the entire list (less than half).
\nSo, I wonder how species in the train data is chosen.

\n

Can you provide us more information about that?

\n

Reference

\n\n

\"primary-list\"
\nFig1. Comparison of train data vs Hawaiian bird species list

", "messages": [ { "id": "1730165", "postDate": "03/21/2022 01:39:39", "content": "

@stefankahl @tomdenton @amandanavine

\n

I have a question to the host.

\n

I downloaded the list of bird species shown in Hawaii[1], and compared it to species in train data.
\nFrom the data, I find out most of the species are in the list, but some are not in the list(Fig. 1).
\nBesides, the species in the train dataset doesn't cover the entire list (less than half).
\nSo, I wonder how species in the train data is chosen.

\n

Can you provide us more information about that?

\n

Reference

\n\n

\"primary-list\"
\nFig1. Comparison of train data vs Hawaiian bird species list

", "rawMarkdown": "stefankahl @tomdenton @amandanavine \n\nI have a question to the host.\n\nI downloaded the list of bird species shown in Hawaii[1], and compared it to species in train data.\nFrom the data, I find out most of the species are in the list, but some are not in the list(Fig. 1).\nBesides, the species in the train dataset doesn't cover the entire list (less than half).\nSo, I wonder how species in the train data is chosen.\n\nCan you provide us more information about that?\n\n## Reference\n\n* [1] http://hbs.bishopmuseum.org/birds/rlp-monograph/PrimaryChecklist.htm\n\n\"primary-list\"\n*Fig1. Comparison of train data vs Hawaiian bird species list*", "votes": null }, { "id": "1730415", "postDate": "03/21/2022 08:19:46", "content": "

The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license. Target species (aka scored species) are those which had expert labels in our test dataset. Some native Hawaiian species might be missing just because there were no appropriate training recordings for these species.

", "rawMarkdown": "The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license. Target species (aka scored species) are those which had expert labels in our test dataset. Some native Hawaiian species might be missing just because there were no appropriate training recordings for these species.", "votes": null }, { "id": "1730440", "postDate": "03/21/2022 08:53:00", "content": "

Thank you for the reply.

\n
\n

The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license.

\n
\n

Based on the above rule, I think a good number of species are missing in the competition dataset.
\nFor example, American Avocet(58), American Bittern(85), Baird's Sandpiper(90) are both satisfies these conditions (the number in parenthesis is the number of recording on Xeno-canto). Besides, they seems to have sufficient number of recordings. Are they just missed or are there any other reasons to be excluded?

\n

c.f.) I compared species in BirdCLEF 2022 dataset and those in Pyle's checklist. There are 161 species missed in this BirdCLEF 2022 dataset.

\n

https://www.kaggle.com/code/tatamikenn/eda-compare-to-the-known-hawaiian-bird-species?scriptVersionId=90803320

", "rawMarkdown": "Thank you for the reply.\n\n> The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license. \n\nBased on the above rule, I think a good number of species are missing in the competition dataset.\nFor example, American Avocet(58), American Bittern(85), Baird's Sandpiper(90) are both satisfies these conditions (the number in parenthesis is the number of recording on Xeno-canto). Besides, they seems to have sufficient number of recordings. Are they just missed or are there any other reasons to be excluded?\n\nc.f.) I compared species in BirdCLEF 2022 dataset and those in Pyle's checklist. There are 161 species missed in this BirdCLEF 2022 dataset.\n\nhttps://www.kaggle.com/code/tatamikenn/eda-compare-to-the-known-hawaiian-bird-species?scriptVersionId=90803320", "votes": null }, { "id": "1732313", "postDate": "03/23/2022 08:26:50", "content": "

It's totally possible that some species are missing. But the reason is solely our automated selection process. Yet, the only species that matter are the \"scored species\" and the 10 endangered Hawaiian species, all of which were manually added to the set. Other species have been included to compensate the lack of sufficient training data (many of the scored species only have very few samples). Feel free to add species from previous editions or entirely new ones (make sure to respect the Xeno-canto terms of use) if you think that it might help you train a better classifier.

", "rawMarkdown": "It's totally possible that some species are missing. But the reason is solely our automated selection process. Yet, the only species that matter are the \"scored species\" and the 10 endangered Hawaiian species, all of which were manually added to the set. Other species have been included to compensate the lack of sufficient training data (many of the scored species only have very few samples). Feel free to add species from previous editions or entirely new ones (make sure to respect the Xeno-canto terms of use) if you think that it might help you train a better classifier.", "votes": null }, { "id": "1732353", "postDate": "03/23/2022 09:40:23", "content": "

I see. Thank you for the reply.

", "rawMarkdown": "I see. Thank you for the reply.", "votes": null } ], "comments": [ { "id": 1730415, "author_name": "stefankahl", "author_url": "", "post_date": "03/21/2022 08:19:46", "content": "

The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license. Target species (aka scored species) are those which had expert labels in our test dataset. Some native Hawaiian species might be missing just because there were no appropriate training recordings for these species.

", "votes": null, "replies": [ { "id": 1730440, "author_name": "tatamikenn", "author_url": "", "post_date": "03/21/2022 08:53:00", "content": "

Thank you for the reply.

\n
\n

The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license.

\n
\n

Based on the above rule, I think a good number of species are missing in the competition dataset.
\nFor example, American Avocet(58), American Bittern(85), Baird's Sandpiper(90) are both satisfies these conditions (the number in parenthesis is the number of recording on Xeno-canto). Besides, they seems to have sufficient number of recordings. Are they just missed or are there any other reasons to be excluded?

\n

c.f.) I compared species in BirdCLEF 2022 dataset and those in Pyle's checklist. There are 161 species missed in this BirdCLEF 2022 dataset.

\n

https://www.kaggle.com/code/tatamikenn/eda-compare-to-the-known-hawaiian-bird-species?scriptVersionId=90803320

", "votes": null, "replies": [] }, { "id": 1732313, "author_name": "stefankahl", "author_url": "", "post_date": "03/23/2022 08:26:50", "content": "

It's totally possible that some species are missing. But the reason is solely our automated selection process. Yet, the only species that matter are the \"scored species\" and the 10 endangered Hawaiian species, all of which were manually added to the set. Other species have been included to compensate the lack of sufficient training data (many of the scored species only have very few samples). Feel free to add species from previous editions or entirely new ones (make sure to respect the Xeno-canto terms of use) if you think that it might help you train a better classifier.

", "votes": null, "replies": [] }, { "id": 1732353, "author_name": "tatamikenn", "author_url": "", "post_date": "03/23/2022 09:40:23", "content": "

I see. Thank you for the reply.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1730165": "stefankahl @tomdenton @amandanavine \n\nI have a question to the host.\n\nI downloaded the list of bird species shown in Hawaii[1], and compared it to species in train data.\nFrom the data, I find out most of the species are in the list, but some are not in the list(Fig. 1).\nBesides, the species in the train dataset doesn't cover the entire list (less than half).\nSo, I wonder how species in the train data is chosen.\n\nCan you provide us more information about that?\n\n## Reference\n\n* [1] http://hbs.bishopmuseum.org/birds/rlp-monograph/PrimaryChecklist.htm\n\n\"primary-list\"\n*Fig1. Comparison of train data vs Hawaiian bird species list*", "1730415": "The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license. Target species (aka scored species) are those which had expert labels in our test dataset. Some native Hawaiian species might be missing just because there were no appropriate training recordings for these species.", "1730440": "Thank you for the reply.\n\n> The training data includes species that a) have at least one eBird observation on the chain of islands of Hawaii and b) have at least one recording on Xeno-canto with appropriate license. \n\nBased on the above rule, I think a good number of species are missing in the competition dataset.\nFor example, American Avocet(58), American Bittern(85), Baird's Sandpiper(90) are both satisfies these conditions (the number in parenthesis is the number of recording on Xeno-canto). Besides, they seems to have sufficient number of recordings. Are they just missed or are there any other reasons to be excluded?\n\nc.f.) I compared species in BirdCLEF 2022 dataset and those in Pyle's checklist. There are 161 species missed in this BirdCLEF 2022 dataset.\n\nhttps://www.kaggle.com/code/tatamikenn/eda-compare-to-the-known-hawaiian-bird-species?scriptVersionId=90803320", "1732313": "It's totally possible that some species are missing. But the reason is solely our automated selection process. Yet, the only species that matter are the \"scored species\" and the 10 endangered Hawaiian species, all of which were manually added to the set. Other species have been included to compensate the lack of sufficient training data (many of the scored species only have very few samples). Feel free to add species from previous editions or entirely new ones (make sure to respect the Xeno-canto terms of use) if you think that it might help you train a better classifier.", "1732353": "I see. Thank you for the reply." }, "source": "meta" }