Hi all! I'm so happy to see that there are so many of you participating. I wanted to make a few suggestions for things that might (?) help improve performance:
\n\nALL iNaturalist 2017/2018 data is allowed, not just the iNat Idaho subset I provided. It might be possible to leverage visual similarities to other classes in the taxonomy (ones not found in Idaho, for example) to augment the training data, in order to do better in Idaho.
We've provided some synthetic data for some of the classes (the AirSim dataset) on github, this might be useful.
I wanted to clarify that unsupervised methods, such as embeddings, clustering, or background subtraction/foreground detection per-location, can be used on the test set. The stipulation is that you CAN NOT do any annotations of the test set. So if you do use a clustering or embedding method you must match clusters to the target classes automatically, without human supervision.
No promises that any of these ideas will help, just a few things I thought of to try :)
", "messages": [ { "id": "523661", "postDate": "04/26/2019 17:23:19", "content": "Hi all! I'm so happy to see that there are so many of you participating. I wanted to make a few suggestions for things that might (?) help improve performance:
\n\nALL iNaturalist 2017/2018 data is allowed, not just the iNat Idaho subset I provided. It might be possible to leverage visual similarities to other classes in the taxonomy (ones not found in Idaho, for example) to augment the training data, in order to do better in Idaho.
We've provided some synthetic data for some of the classes (the AirSim dataset) on github, this might be useful.
I wanted to clarify that unsupervised methods, such as embeddings, clustering, or background subtraction/foreground detection per-location, can be used on the test set. The stipulation is that you CAN NOT do any annotations of the test set. So if you do use a clustering or embedding method you must match clusters to the target classes automatically, without human supervision.
No promises that any of these ideas will help, just a few things I thought of to try :)
", "rawMarkdown": "Hi all! I'm so happy to see that there are so many of you participating. I wanted to make a few suggestions for things that might (?) help improve performance:\n\n1. ALL iNaturalist 2017/2018 data is allowed, not just the iNat Idaho subset I provided. It might be possible to leverage visual similarities to other classes in the taxonomy (ones not found in Idaho, for example) to augment the training data, in order to do better in Idaho.\n\n2. We've provided some synthetic data for some of the classes (the AirSim dataset) on github, this might be useful.\n\n3. I wanted to clarify that unsupervised methods, such as embeddings, clustering, or background subtraction/foreground detection per-location, can be used on the test set. The stipulation is that you CAN NOT do any annotations of the test set. So if you do use a clustering or embedding method you must match clusters to the target classes automatically, without human supervision.\n\nNo promises that any of these ideas will help, just a few things I thought of to try :)", "votes": null }, { "id": "524604", "postDate": "04/29/2019 07:05:48", "content": "Are you going to add a link to the whole iNat dataset in GitHub? How big it is?
", "rawMarkdown": "Are you going to add a link to the whole iNat dataset in GitHub? How big it is?", "votes": null }, { "id": "524793", "postDate": "04/29/2019 14:33:08", "content": "You can access the iNaturalist data directly from the iNaturalist 2017 and 2018 competition pages. The data is large.
", "rawMarkdown": "You can access the iNaturalist data directly from the iNaturalist [2017](https://github.com/visipedia/inat_comp/tree/master/2017) and [2018](https://github.com/visipedia/inat_comp) competition pages. The data is large.", "votes": null }, { "id": "528993", "postDate": "05/09/2019 02:13:35", "content": "Can I use the ID(image name) to clustering for the test set.
", "rawMarkdown": "Can I use the ID(image name) to clustering for the test set.", "votes": null }, { "id": "529055", "postDate": "05/09/2019 05:02:42", "content": "No. The ID should be randomized and should not contain relevant information. Just in case I made a mistake and there is relevant information in the ID, I'm excluding it. If you had new images from a new region you wouldn't have an \"ID\" to cluster them by, and we're looking for solutions that will be helpful in the real world.
", "rawMarkdown": "No. The ID should be randomized and should not contain relevant information. Just in case I made a mistake and there is relevant information in the ID, I'm excluding it. If you had new images from a new region you wouldn't have an \"ID\" to cluster them by, and we're looking for solutions that will be helpful in the real world.", "votes": null } ], "comments": [ { "id": 524604, "author_name": "andreapi", "author_url": "", "post_date": "04/29/2019 07:05:48", "content": "Are you going to add a link to the whole iNat dataset in GitHub? How big it is?
", "votes": null, "replies": [ { "id": 524793, "author_name": "sbeery", "author_url": "", "post_date": "04/29/2019 14:33:08", "content": "You can access the iNaturalist data directly from the iNaturalist 2017 and 2018 competition pages. The data is large.
", "votes": null, "replies": [] } ] }, { "id": 528993, "author_name": "fengepsilon", "author_url": "", "post_date": "05/09/2019 02:13:35", "content": "Can I use the ID(image name) to clustering for the test set.
", "votes": null, "replies": [ { "id": 529055, "author_name": "sbeery", "author_url": "", "post_date": "05/09/2019 05:02:42", "content": "No. The ID should be randomized and should not contain relevant information. Just in case I made a mistake and there is relevant information in the ID, I'm excluding it. If you had new images from a new region you wouldn't have an \"ID\" to cluster them by, and we're looking for solutions that will be helpful in the real world.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "523661": "Hi all! I'm so happy to see that there are so many of you participating. I wanted to make a few suggestions for things that might (?) help improve performance:\n\n1. ALL iNaturalist 2017/2018 data is allowed, not just the iNat Idaho subset I provided. It might be possible to leverage visual similarities to other classes in the taxonomy (ones not found in Idaho, for example) to augment the training data, in order to do better in Idaho.\n\n2. We've provided some synthetic data for some of the classes (the AirSim dataset) on github, this might be useful.\n\n3. I wanted to clarify that unsupervised methods, such as embeddings, clustering, or background subtraction/foreground detection per-location, can be used on the test set. The stipulation is that you CAN NOT do any annotations of the test set. So if you do use a clustering or embedding method you must match clusters to the target classes automatically, without human supervision.\n\nNo promises that any of these ideas will help, just a few things I thought of to try :)", "524604": "Are you going to add a link to the whole iNat dataset in GitHub? How big it is?", "524793": "You can access the iNaturalist data directly from the iNaturalist [2017](https://github.com/visipedia/inat_comp/tree/master/2017) and [2018](https://github.com/visipedia/inat_comp) competition pages. The data is large.", "528993": "Can I use the ID(image name) to clustering for the test set.", "529055": "No. The ID should be randomized and should not contain relevant information. Just in case I made a mistake and there is relevant information in the ID, I'm excluding it. If you had new images from a new region you wouldn't have an \"ID\" to cluster them by, and we're looking for solutions that will be helpful in the real world." }, "source": "meta" }