Hello everyone.
\nI've searched in existing discussions and found some people seemed to be encountering the same issue as I am encountering myself, but did not find any explicit discussion about it.
\nSo here is my problem : I am wondering if all the labels are correct or not. For instance, I checked some images belonging to the \"healthy\" class (class 4). I found three images labeled \"healthy\" which seem suspicious :
\n1001723730.jpg
\n1012902365.jpg
\n1022932733.jpg
\nOf course, I've never cultivated cassava myself, but these yellow leaves do not seem so healthy… Except if these are old leaves ? These three examples belong to the first 132 lines of the train.csv file, so I assume there are many more such examples.
\nThank you very much if you have anything about this !
Hello everyone.
\nI've searched in existing discussions and found some people seemed to be encountering the same issue as I am encountering myself, but did not find any explicit discussion about it.
\nSo here is my problem : I am wondering if all the labels are correct or not. For instance, I checked some images belonging to the \"healthy\" class (class 4). I found three images labeled \"healthy\" which seem suspicious :
\n1001723730.jpg
\n1012902365.jpg
\n1022932733.jpg
\nOf course, I've never cultivated cassava myself, but these yellow leaves do not seem so healthy… Except if these are old leaves ? These three examples belong to the first 132 lines of the train.csv file, so I assume there are many more such examples.
\nThank you very much if you have anything about this !
There are multiple such images. There are also images of cassava potatoes instead of leaves. I have made a notebook to find these with cosine-similarity: https://www.kaggle.com/aliabdin1/find-similar-images-with-cosine-similarity
\nThe competition data is noisy and one challenge of this competition is to remove or make the noise as much as irrelevant as possible.
", "rawMarkdown": "There are multiple such images. There are also images of cassava potatoes instead of leaves. I have made a notebook to find these with cosine-similarity: https://www.kaggle.com/aliabdin1/find-similar-images-with-cosine-similarity\n\nThe competition data is noisy and one challenge of this competition is to remove or make the noise as much as irrelevant as possible.", "votes": null }, { "id": "1147919", "postDate": "01/10/2021 19:59:13", "content": "Thank you very much for your answer.
", "rawMarkdown": "Thank you very much for your answer.", "votes": null }, { "id": "1147927", "postDate": "01/10/2021 20:04:04", "content": "Im glad to help, are you planning to remove them and see if your CV/LB increases?
", "rawMarkdown": "Im glad to help, are you planning to remove them and see if your CV/LB increases?", "votes": null }, { "id": "1147964", "postDate": "01/10/2021 20:26:46", "content": "No that would take too much time. I used this dataset to learn more about transfer learning. I used CNNs before on MNIST digits recognition and Pneumonia X-Ray datasets but didn't experiment transfer learning for these tasks.
", "rawMarkdown": "No that would take too much time. I used this dataset to learn more about transfer learning. I used CNNs before on MNIST digits recognition and Pneumonia X-Ray datasets but didn't experiment transfer learning for these tasks.", "votes": null } ], "comments": [ { "id": 1132892, "author_name": "aliabdin1", "author_url": "", "post_date": "12/30/2020 19:04:57", "content": "There are multiple such images. There are also images of cassava potatoes instead of leaves. I have made a notebook to find these with cosine-similarity: https://www.kaggle.com/aliabdin1/find-similar-images-with-cosine-similarity
\nThe competition data is noisy and one challenge of this competition is to remove or make the noise as much as irrelevant as possible.
", "votes": null, "replies": [ { "id": 1147919, "author_name": "florian12", "author_url": "", "post_date": "01/10/2021 19:59:13", "content": "Thank you very much for your answer.
", "votes": null, "replies": [] }, { "id": 1147927, "author_name": "aliabdin1", "author_url": "", "post_date": "01/10/2021 20:04:04", "content": "Im glad to help, are you planning to remove them and see if your CV/LB increases?
", "votes": null, "replies": [] }, { "id": 1147964, "author_name": "florian12", "author_url": "", "post_date": "01/10/2021 20:26:46", "content": "No that would take too much time. I used this dataset to learn more about transfer learning. I used CNNs before on MNIST digits recognition and Pneumonia X-Ray datasets but didn't experiment transfer learning for these tasks.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1132641": "Hello everyone.\nI've searched in existing discussions and found some people seemed to be encountering the same issue as I am encountering myself, but did not find any explicit discussion about it.\nSo here is my problem : I am wondering if all the labels are correct or not. For instance, I checked some images belonging to the \"healthy\" class (class 4). I found three images labeled \"healthy\" which seem suspicious :\n1001723730.jpg\n1012902365.jpg\n1022932733.jpg\nOf course, I've never cultivated cassava myself, but these yellow leaves do not seem so healthy… Except if these are old leaves ? These three examples belong to the first 132 lines of the train.csv file, so I assume there are many more such examples.\nThank you very much if you have anything about this !", "1132892": "There are multiple such images. There are also images of cassava potatoes instead of leaves. I have made a notebook to find these with cosine-similarity: https://www.kaggle.com/aliabdin1/find-similar-images-with-cosine-similarity\n\nThe competition data is noisy and one challenge of this competition is to remove or make the noise as much as irrelevant as possible.", "1147919": "Thank you very much for your answer.", "1147927": "Im glad to help, are you planning to remove them and see if your CV/LB increases?", "1147964": "No that would take too much time. I used this dataset to learn more about transfer learning. I used CNNs before on MNIST digits recognition and Pneumonia X-Ray datasets but didn't experiment transfer learning for these tasks." }, "source": "meta" }