Welcome to the Humpback Whale Identification Challenge!
\n\nHelp us master automated image recognition of individual whales! ID of individual whales is an incredibly powerful tool for science. And, it’s really cool to be able to track whales through their travels and migrations :). This competition is built on one of the world’s largest individual animal ID catalogs, data from citizen science contributions to Happywhale.com and collaborating researchers. Building a global database of individual whales is helping us understand population trends and conservation issues, increasingly pressing questions as we use our oceans ever more intensively.
\n\nOne-shot image recognition has come far but can still be improved — note a recent publication by Weideman et al (2017) on algorithms trained for dolphin and whale individual ID.
\n\nThank you for participating — we welcome your efforts
", "messages": [ { "id": "432652", "postDate": "12/04/2018 06:22:34", "content": "Welcome to the Humpback Whale Identification Challenge!
\n\nHelp us master automated image recognition of individual whales! ID of individual whales is an incredibly powerful tool for science. And, it’s really cool to be able to track whales through their travels and migrations :). This competition is built on one of the world’s largest individual animal ID catalogs, data from citizen science contributions to Happywhale.com and collaborating researchers. Building a global database of individual whales is helping us understand population trends and conservation issues, increasingly pressing questions as we use our oceans ever more intensively.
\n\nOne-shot image recognition has come far but can still be improved — note a recent publication by Weideman et al (2017) on algorithms trained for dolphin and whale individual ID.
\n\nThank you for participating — we welcome your efforts
", "rawMarkdown": "Welcome to the Humpback Whale Identification Challenge!\n\nHelp us master automated image recognition of individual whales! ID of individual whales is an incredibly powerful tool for science. And, it’s really cool to be able to track whales through their travels and migrations :). This competition is built on one of the world’s largest individual animal ID catalogs, data from citizen science contributions to [**Happywhale.com**](https://happywhale.com) and collaborating researchers. Building a global database of individual whales is helping us understand population trends and conservation issues, increasingly pressing questions as we use our oceans ever more intensively.\n\nOne-shot image recognition has come far but can still be improved — note a [**recent publication by Weideman et al (2017)**](https://arxiv.org/pdf/1708.07785.pdf) on algorithms trained for dolphin and whale individual ID.\n\nThank you for participating — we welcome your efforts", "votes": null }, { "id": "434295", "postDate": "12/06/2018 07:00:45", "content": "Thank you for a warm welcome ;)
", "rawMarkdown": "Thank you for a warm welcome ;)", "votes": null }, { "id": "436435", "postDate": "12/10/2018 09:59:27", "content": "In the paper predicting the trailing edge of the fluke is mentioned using segmentation. Would you have a dataset with trailing edge markings available? That would be of great help I feel.
\n\nDoes anyone know where such data can be obtained?
", "rawMarkdown": "In the paper predicting the trailing edge of the fluke is mentioned using segmentation. Would you have a dataset with trailing edge markings available? That would be of great help I feel.\n\nDoes anyone know where such data can be obtained?", "votes": null }, { "id": "436662", "postDate": "12/10/2018 17:58:45", "content": "The paper tells you how to calculate it. Use concentric circles to sample the pixel edge and level the segments by rotation around the concentric circle center.
", "rawMarkdown": "The paper tells you how to calculate it. Use concentric circles to sample the pixel edge and level the segments by rotation around the concentric circle center.", "votes": null }, { "id": "439492", "postDate": "12/15/2018 15:35:49", "content": "Thank you very much for your post! I am excited to develop a great model.
", "rawMarkdown": "Thank you very much for your post! I am excited to develop a great model.", "votes": null }, { "id": "440578", "postDate": "12/17/2018 18:13:48", "content": "I think he means calculating the co-ordinate vectors marking the trailing edge which they calculate as follows : \n > Given an image, a fully-convolutional neural network\n> (FCNN) [16] outputs the probability that each pixel is part\n> of the trailing edge. Anchor points are computed, and a\n> shortest-path algorithm selects pixels based on costs determined by a combination of the FCNN and > > > image gradients
\n\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
", "rawMarkdown": "I think he means calculating the co-ordinate vectors marking the trailing edge which they calculate as follows : \n > Given an image, a fully-convolutional neural network\n> (FCNN) [16] outputs the probability that each pixel is part\n> of the trailing edge. Anchor points are computed, and a\n> shortest-path algorithm selects pixels based on costs determined by a combination of the FCNN and > > > image gradients\n\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf", "votes": null }, { "id": "440580", "postDate": "12/17/2018 18:14:16", "content": "It would indeed be great if we could get the trailing edge markings as a dataset.
", "rawMarkdown": "It would indeed be great if we could get the trailing edge markings as a dataset.", "votes": null }, { "id": "441220", "postDate": "12/18/2018 12:31:10", "content": "Regrets that we don't have such a dataset immediately available. If I can find something, I will share
", "rawMarkdown": "Regrets that we don't have such a dataset immediately available. If I can find something, I will share", "votes": null }, { "id": "443043", "postDate": "12/20/2018 23:21:53", "content": "Many of you have found Martin Piotte's solution to the playground a while back. Here is a repo of the code reworked by me to be set up in a way that is directly usable by our ID system. In the end, we would love to see a command-line solution that has a train function and an identification function. The identification function takes a directory name and performs an ID on all images inside that directory. In our production server, the user uploads a set of images to a temporary folder and the identification is performed on that folder. So have a look-see and good luck all.
", "rawMarkdown": "Many of you have found [Martin Piotte's](https://www.kaggle.com/martinpiotte/whale-recognition-model-with-score-0-78563) solution to the playground a while back. Here is a [repo](https://github.com/animalus/whalerec) of the code reworked by me to be set up in a way that is directly usable by our ID system. In the end, we would love to see a command-line solution that has a train function and an identification function. The identification function takes a directory name and performs an ID on all images inside that directory. In our production server, the user uploads a set of images to a temporary folder and the identification is performed on that folder. So have a look-see and good luck all.", "votes": null }, { "id": "443169", "postDate": "12/21/2018 06:20:12", "content": "Thanks for your warm welcome!
", "rawMarkdown": "Thanks for your warm welcome!", "votes": null }, { "id": "455311", "postDate": "01/13/2019 15:15:18", "content": "May be I am just lazy, but having algorithm from the paper you meantioned open-sourced, would definitely help participants :) Re-implementing it from paper you mentioned (Weideman et al (2017) ) is not an easy task
", "rawMarkdown": "May be I am just lazy, but having algorithm from the paper you meantioned open-sourced, would definitely help participants :) Re-implementing it from paper you mentioned (Weideman et al (2017) ) is not an easy task", "votes": null }, { "id": "457462", "postDate": "01/17/2019 13:33:20", "content": "Did you run any tests of this reworked code/model for the present competition data? LB score?
", "rawMarkdown": "Did you run any tests of this reworked code/model for the present competition data? LB score?", "votes": null }, { "id": "460534", "postDate": "01/23/2019 22:55:07", "content": "Have you guys at happywhale.com ever considered collecting whale fluke data using video? Maybe not for user submitted inference (data bandwidth and compute requirements would probably increase significantly) but for researchers collecting training data it provides a lot of extra capabilities.
\n\nThanks Paul. Interesting thought. Main drawbacks are the processing complexities. If the winning algorithms of this kaggle contest prove successfully implementable and score as high as the leaderboard suggests, we'll be well set with IDs from photos, which are cheap and non-invasive and very accessible / available from citizen scientists.
", "rawMarkdown": "Thanks Paul. Interesting thought. Main drawbacks are the processing complexities. If the winning algorithms of this kaggle contest prove successfully implementable and score as high as the leaderboard suggests, we'll be well set with IDs from photos, which are cheap and non-invasive and very accessible / available from citizen scientists.", "votes": null }, { "id": "475506", "postDate": "02/20/2019 20:28:37", "content": "Next time use a good infrared camera, will make the job easier (in my opinion).\nBy the way, i did my best, 0.8 score.\nI really enjoyed this challenge, thank you very much @tedcheese and happywhale for your job and for let us to contribute.
", "rawMarkdown": "Next time use a good infrared camera, will make the job easier (in my opinion).\nBy the way, i did my best, **0.8 score**.\nI really enjoyed this challenge, thank you very much @tedcheese and [happywhale](https://happywhale.com/) for your job and for let us to contribute.", "votes": null }, { "id": "480559", "postDate": "02/28/2019 10:41:36", "content": "Hi Team
\n\nThis was a great competition enjoyed researching and learning so far. Did my best to help recognize the whales. Learned a lot of stuff here with some are comparing the tails and some are using pretrained weights,
", "rawMarkdown": "Hi Team\n\nThis was a great competition enjoyed researching and learning so far. Did my best to help recognize the whales. Learned a lot of stuff here with some are comparing the tails and some are using pretrained weights,", "votes": null }, { "id": "483607", "postDate": "03/04/2019 21:32:15", "content": "hi team,
\n\nshould i need to use deep learning for this challenge?? I am beginner in this challenge
", "rawMarkdown": "hi team, \n\nshould i need to use deep learning for this challenge?? I am beginner in this challenge", "votes": null }, { "id": "1501366", "postDate": "09/03/2021 06:29:08", "content": "Hi Ted, I would like to use a few samples from the dataset in a book I am writing. What is the license on the dataset usage? I couldn't find it here.
", "rawMarkdown": "Hi Ted, I would like to use a few samples from the dataset in a book I am writing. What is the license on the dataset usage? I couldn't find it here.", "votes": null }, { "id": "1502257", "postDate": "09/04/2021 04:44:15", "content": "Hi. Thank you for asking. Image use varies by contributor. If you give me a few samples of images you would like to use I can look up the license on each. Best to contact me directly - ted@happywhale.com
", "rawMarkdown": "Hi. Thank you for asking. Image use varies by contributor. If you give me a few samples of images you would like to use I can look up the license on each. Best to contact me directly - ted@happywhale.com", "votes": null } ], "comments": [ { "id": 1501366, "author_name": "naresh", "author_url": "", "post_date": "09/03/2021 06:29:08", "content": "Hi Ted, I would like to use a few samples from the dataset in a book I am writing. What is the license on the dataset usage? I couldn't find it here.
", "votes": null, "replies": [ { "id": 1502257, "author_name": "tedcheese", "author_url": "", "post_date": "09/04/2021 04:44:15", "content": "Hi. Thank you for asking. Image use varies by contributor. If you give me a few samples of images you would like to use I can look up the license on each. Best to contact me directly - ted@happywhale.com
", "votes": null, "replies": [] } ] }, { "id": 434295, "author_name": "codealist", "author_url": "", "post_date": "12/06/2018 07:00:45", "content": "Thank you for a warm welcome ;)
", "votes": null, "replies": [] }, { "id": 436435, "author_name": "radek1", "author_url": "", "post_date": "12/10/2018 09:59:27", "content": "In the paper predicting the trailing edge of the fluke is mentioned using segmentation. Would you have a dataset with trailing edge markings available? That would be of great help I feel.
\n\nDoes anyone know where such data can be obtained?
", "votes": null, "replies": [ { "id": 436662, "author_name": "badtyprr", "author_url": "", "post_date": "12/10/2018 17:58:45", "content": "The paper tells you how to calculate it. Use concentric circles to sample the pixel edge and level the segments by rotation around the concentric circle center.
", "votes": null, "replies": [] }, { "id": 440578, "author_name": "cheesydionysus", "author_url": "", "post_date": "12/17/2018 18:13:48", "content": "I think he means calculating the co-ordinate vectors marking the trailing edge which they calculate as follows : \n > Given an image, a fully-convolutional neural network\n> (FCNN) [16] outputs the probability that each pixel is part\n> of the trailing edge. Anchor points are computed, and a\n> shortest-path algorithm selects pixels based on costs determined by a combination of the FCNN and > > > image gradients
\n\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
", "votes": null, "replies": [] }, { "id": 440580, "author_name": "cheesydionysus", "author_url": "", "post_date": "12/17/2018 18:14:16", "content": "It would indeed be great if we could get the trailing edge markings as a dataset.
", "votes": null, "replies": [] }, { "id": 441220, "author_name": "tedcheese", "author_url": "", "post_date": "12/18/2018 12:31:10", "content": "Regrets that we don't have such a dataset immediately available. If I can find something, I will share
", "votes": null, "replies": [] }, { "id": 455311, "author_name": "oldufo", "author_url": "", "post_date": "01/13/2019 15:15:18", "content": "May be I am just lazy, but having algorithm from the paper you meantioned open-sourced, would definitely help participants :) Re-implementing it from paper you mentioned (Weideman et al (2017) ) is not an easy task
", "votes": null, "replies": [] } ] }, { "id": 439492, "author_name": "sangwookchn", "author_url": "", "post_date": "12/15/2018 15:35:49", "content": "Thank you very much for your post! I am excited to develop a great model.
", "votes": null, "replies": [] }, { "id": 443043, "author_name": "crowmagnumb", "author_url": "", "post_date": "12/20/2018 23:21:53", "content": "Many of you have found Martin Piotte's solution to the playground a while back. Here is a repo of the code reworked by me to be set up in a way that is directly usable by our ID system. In the end, we would love to see a command-line solution that has a train function and an identification function. The identification function takes a directory name and performs an ID on all images inside that directory. In our production server, the user uploads a set of images to a temporary folder and the identification is performed on that folder. So have a look-see and good luck all.
", "votes": null, "replies": [ { "id": 457462, "author_name": "pinullmezon", "author_url": "", "post_date": "01/17/2019 13:33:20", "content": "Did you run any tests of this reworked code/model for the present competition data? LB score?
", "votes": null, "replies": [] } ] }, { "id": 443169, "author_name": "zhuonanlin383", "author_url": "", "post_date": "12/21/2018 06:20:12", "content": "Thanks for your warm welcome!
", "votes": null, "replies": [] }, { "id": 460534, "author_name": "oewyn000", "author_url": "", "post_date": "01/23/2019 22:55:07", "content": "Have you guys at happywhale.com ever considered collecting whale fluke data using video? Maybe not for user submitted inference (data bandwidth and compute requirements would probably increase significantly) but for researchers collecting training data it provides a lot of extra capabilities.
\n\nThanks Paul. Interesting thought. Main drawbacks are the processing complexities. If the winning algorithms of this kaggle contest prove successfully implementable and score as high as the leaderboard suggests, we'll be well set with IDs from photos, which are cheap and non-invasive and very accessible / available from citizen scientists.
", "votes": null, "replies": [] } ] }, { "id": 475506, "author_name": "jesucristo", "author_url": "", "post_date": "02/20/2019 20:28:37", "content": "Next time use a good infrared camera, will make the job easier (in my opinion).\nBy the way, i did my best, 0.8 score.\nI really enjoyed this challenge, thank you very much @tedcheese and happywhale for your job and for let us to contribute.
", "votes": null, "replies": [] }, { "id": 480559, "author_name": "akhileshrai003", "author_url": "", "post_date": "02/28/2019 10:41:36", "content": "Hi Team
\n\nThis was a great competition enjoyed researching and learning so far. Did my best to help recognize the whales. Learned a lot of stuff here with some are comparing the tails and some are using pretrained weights,
", "votes": null, "replies": [] }, { "id": 483607, "author_name": "mtokons", "author_url": "", "post_date": "03/04/2019 21:32:15", "content": "hi team,
\n\nshould i need to use deep learning for this challenge?? I am beginner in this challenge
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "432652": "Welcome to the Humpback Whale Identification Challenge!\n\nHelp us master automated image recognition of individual whales! ID of individual whales is an incredibly powerful tool for science. And, it’s really cool to be able to track whales through their travels and migrations :). This competition is built on one of the world’s largest individual animal ID catalogs, data from citizen science contributions to [**Happywhale.com**](https://happywhale.com) and collaborating researchers. Building a global database of individual whales is helping us understand population trends and conservation issues, increasingly pressing questions as we use our oceans ever more intensively.\n\nOne-shot image recognition has come far but can still be improved — note a [**recent publication by Weideman et al (2017)**](https://arxiv.org/pdf/1708.07785.pdf) on algorithms trained for dolphin and whale individual ID.\n\nThank you for participating — we welcome your efforts", "434295": "Thank you for a warm welcome ;)", "436435": "In the paper predicting the trailing edge of the fluke is mentioned using segmentation. Would you have a dataset with trailing edge markings available? That would be of great help I feel.\n\nDoes anyone know where such data can be obtained?", "436662": "The paper tells you how to calculate it. Use concentric circles to sample the pixel edge and level the segments by rotation around the concentric circle center.", "439492": "Thank you very much for your post! I am excited to develop a great model.", "440578": "I think he means calculating the co-ordinate vectors marking the trailing edge which they calculate as follows : \n > Given an image, a fully-convolutional neural network\n> (FCNN) [16] outputs the probability that each pixel is part\n> of the trailing edge. Anchor points are computed, and a\n> shortest-path algorithm selects pixels based on costs determined by a combination of the FCNN and > > > image gradients\n\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf", "440580": "It would indeed be great if we could get the trailing edge markings as a dataset.", "441220": "Regrets that we don't have such a dataset immediately available. If I can find something, I will share", "443043": "Many of you have found [Martin Piotte's](https://www.kaggle.com/martinpiotte/whale-recognition-model-with-score-0-78563) solution to the playground a while back. Here is a [repo](https://github.com/animalus/whalerec) of the code reworked by me to be set up in a way that is directly usable by our ID system. In the end, we would love to see a command-line solution that has a train function and an identification function. The identification function takes a directory name and performs an ID on all images inside that directory. In our production server, the user uploads a set of images to a temporary folder and the identification is performed on that folder. So have a look-see and good luck all.", "443169": "Thanks for your warm welcome!", "455311": "May be I am just lazy, but having algorithm from the paper you meantioned open-sourced, would definitely help participants :) Re-implementing it from paper you mentioned (Weideman et al (2017) ) is not an easy task", "457462": "Did you run any tests of this reworked code/model for the present competition data? LB score?", "460534": "Have you guys at happywhale.com ever considered collecting whale fluke data using video? Maybe not for user submitted inference (data bandwidth and compute requirements would probably increase significantly) but for researchers collecting training data it provides a lot of extra capabilities.\n\n * frames can be extracted from a video to get multiple images of the same fluke for identification. -- You don't necessarily have to assign/match it to a whale_id.\n* Cases where occlusions/obstructions make identification difficult, additional frames are available that might have additional information.\n * You can utilize optical flow information between frames to automatically segment video to mask out only the pixels belonging to the whale.", "461459": "Thanks Paul. Interesting thought. Main drawbacks are the processing complexities. If the winning algorithms of this kaggle contest prove successfully implementable and score as high as the leaderboard suggests, we'll be well set with IDs from photos, which are cheap and non-invasive and very accessible / available from citizen scientists.", "475506": "Next time use a good infrared camera, will make the job easier (in my opinion).\nBy the way, i did my best, **0.8 score**.\nI really enjoyed this challenge, thank you very much @tedcheese and [happywhale](https://happywhale.com/) for your job and for let us to contribute.", "480559": "Hi Team\n\nThis was a great competition enjoyed researching and learning so far. Did my best to help recognize the whales. Learned a lot of stuff here with some are comparing the tails and some are using pretrained weights,", "483607": "hi team, \n\nshould i need to use deep learning for this challenge?? I am beginner in this challenge", "1501366": "Hi Ted, I would like to use a few samples from the dataset in a book I am writing. What is the license on the dataset usage? I couldn't find it here.", "1502257": "Hi. Thank you for asking. Image use varies by contributor. If you give me a few samples of images you would like to use I can look up the license on each. Best to contact me directly - ted@happywhale.com" }, "source": "meta" }