There are so many great notebook and resources out here for this competition, so I am pretty sure this kind of analysis has been done before. I think one of the main challenges for this dataset is the low number of samples for most of the individuals.
\nMy analysis showed, that:
\n59.40% of all individuals are represented by 1 sample in the training set.
\n79.23% of all individuals are represented by a maximum of 2 samples in the training set.
\n84.19% of all individuals are represented by a maximum of 3 samples in the training set.
For all deep learning models, this is a pretty low number of samples to train with.
\nThe data is really great to train a species classifier, but moving on to recognizing individuals seems pretty tough. I mean pretty tough to train of course, but also very hard to evaluate, e.g. if you only have 1 sample for more than half of the classes.
While I have been doing a lot of work in general object detection and classification, I am pretty new to the field of individual recognition. Can somebody give me a hint of what is the current SOTA for this kind of problem?
", "messages": [ { "id": "1682833", "postDate": "02/09/2022 12:08:43", "content": "There are so many great notebook and resources out here for this competition, so I am pretty sure this kind of analysis has been done before. I think one of the main challenges for this dataset is the low number of samples for most of the individuals.
\nMy analysis showed, that:
\n59.40% of all individuals are represented by 1 sample in the training set.
\n79.23% of all individuals are represented by a maximum of 2 samples in the training set.
\n84.19% of all individuals are represented by a maximum of 3 samples in the training set.
For all deep learning models, this is a pretty low number of samples to train with.
\nThe data is really great to train a species classifier, but moving on to recognizing individuals seems pretty tough. I mean pretty tough to train of course, but also very hard to evaluate, e.g. if you only have 1 sample for more than half of the classes.
While I have been doing a lot of work in general object detection and classification, I am pretty new to the field of individual recognition. Can somebody give me a hint of what is the current SOTA for this kind of problem?
", "rawMarkdown": "There are so many great notebook and resources out here for this competition, so I am pretty sure this kind of analysis has been done before. I think one of the main challenges for this dataset is the low number of samples for most of the individuals.\n\nMy analysis showed, that:\n**59.40%** of all individuals are represented by 1 sample in the training set.\n**79.23%** of all individuals are represented by a maximum of 2 samples in the training set.\n**84.19%** of all individuals are represented by a maximum of 3 samples in the training set.\n\nFor all deep learning models, this is a pretty low number of samples to train with. \nThe data is really great to train a species classifier, but moving on to recognizing individuals seems pretty tough. I mean pretty tough to train of course, but also very hard to evaluate, e.g. if you only have 1 sample for more than half of the *classes*.\n\nWhile I have been doing a lot of work in general object detection and classification, I am pretty new to the field of individual recognition. Can somebody give me a hint of what is the current SOTA for this kind of problem?", "votes": null }, { "id": "1683353", "postDate": "02/09/2022 18:49:11", "content": "Kaggle has a number of conceptually similar competitions where you can browse winning strategies. The recent ones I found to be most helpful were:
\nI would suggest you to kindly check out the winning solutions from the previous whale challenge. That's really close to this one in terms of data too 🐳
\nSiamese networks or forms of \"one-shot learning\" might apply well here.
\nGood Luck! :)
", "rawMarkdown": "I would suggest you to kindly check out the winning solutions from the previous whale challenge. That's really close to this one in terms of data too 🐳\n\nSiamese networks or forms of \"one-shot learning\" might apply well here. \n\nGood Luck! :)", "votes": null } ], "comments": [ { "id": 1683353, "author_name": "rturley", "author_url": "", "post_date": "02/09/2022 18:49:11", "content": "Kaggle has a number of conceptually similar competitions where you can browse winning strategies. The recent ones I found to be most helpful were:
\nI would suggest you to kindly check out the winning solutions from the previous whale challenge. That's really close to this one in terms of data too 🐳
\nSiamese networks or forms of \"one-shot learning\" might apply well here.
\nGood Luck! :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1682833": "There are so many great notebook and resources out here for this competition, so I am pretty sure this kind of analysis has been done before. I think one of the main challenges for this dataset is the low number of samples for most of the individuals.\n\nMy analysis showed, that:\n**59.40%** of all individuals are represented by 1 sample in the training set.\n**79.23%** of all individuals are represented by a maximum of 2 samples in the training set.\n**84.19%** of all individuals are represented by a maximum of 3 samples in the training set.\n\nFor all deep learning models, this is a pretty low number of samples to train with. \nThe data is really great to train a species classifier, but moving on to recognizing individuals seems pretty tough. I mean pretty tough to train of course, but also very hard to evaluate, e.g. if you only have 1 sample for more than half of the *classes*.\n\nWhile I have been doing a lot of work in general object detection and classification, I am pretty new to the field of individual recognition. Can somebody give me a hint of what is the current SOTA for this kind of problem?", "1683353": "Kaggle has a number of conceptually similar competitions where you can browse winning strategies. The recent ones I found to be most helpful were: \n- [Humpback Whale Identification](https://www.kaggle.com/c/humpback-whale-identification/) (3000+ whales)\n- [Google Landmark Recognition 2021](https://www.kaggle.com/c/landmark-recognition-2021/) (81k landmarks)\n- [Hotel-ID to Combat Human Trafficking 2021](https://www.kaggle.com/c/hotel-id-2021-fgvc8) (7700 hotels)", "1686055": "I would suggest you to kindly check out the winning solutions from the previous whale challenge. That's really close to this one in terms of data too 🐳\n\nSiamese networks or forms of \"one-shot learning\" might apply well here. \n\nGood Luck! :)" }, "source": "meta" }