Hi,
\nI am a little confused with how the bounding boxes are labeled. While we have a study level CSV file (train_study_level.csv) which provides 4 different categories for each study, we also have image-level labeling which provides either none or opacity for bounding boxes. For images that have more than one opacity, should we assume all opacity has the same category provided by train_study_level.csv or something else?
\nThe reason I am asking this is that I have noticed in https://www.kaggle.com/c/siim-covid19-detection/overview/evaluation, you can provide different labels for each opacity.
Hi,
\nI am a little confused with how the bounding boxes are labeled. While we have a study level CSV file (train_study_level.csv) which provides 4 different categories for each study, we also have image-level labeling which provides either none or opacity for bounding boxes. For images that have more than one opacity, should we assume all opacity has the same category provided by train_study_level.csv or something else?
\nThe reason I am asking this is that I have noticed in https://www.kaggle.com/c/siim-covid19-detection/overview/evaluation, you can provide different labels for each opacity.
As far as I understand it, you're making a study level prediction for the image in the study. Then you are making a separate prediction of the opacity labels on the image.
\nSo for the same image you would:
\n1) Look at its study id and make a prediction out of the 4 categories. (1 prediction per image)
\n2) Look at its image id and make object detections of opacities. Here you only have 1 category -- opacity. (Could have multiple predictions per image)
There might well be some correlation between the number of opacities and what category the study is in, but you are certainly not predicting a category every time you detect an opacity.
", "rawMarkdown": "As far as I understand it, you're making a study level prediction for the image in the study. Then you are making a separate prediction of the opacity labels on the image. \nSo for the same image you would:\n1) Look at its study id and make a prediction out of the 4 categories. (1 prediction per image)\n2) Look at its image id and make object detections of opacities. Here you only have 1 category -- opacity. (Could have multiple predictions per image)\n\nThere might well be some correlation between the number of opacities and what category the study is in, but you are certainly not predicting a category every time you detect an opacity.", "votes": null }, { "id": "1398704", "postDate": "07/24/2021 12:15:12", "content": "My understanding is it is possible for study level to have more than one label, as each image have more than one bbox and also some study also refer to more than one images. Please correct me if I am wrong
", "rawMarkdown": "My understanding is it is possible for study level to have more than one label, as each image have more than one bbox and also some study also refer to more than one images. Please correct me if I am wrong", "votes": null } ], "comments": [ { "id": 1398222, "author_name": "diningeachox", "author_url": "", "post_date": "07/23/2021 21:34:14", "content": "As far as I understand it, you're making a study level prediction for the image in the study. Then you are making a separate prediction of the opacity labels on the image.
\nSo for the same image you would:
\n1) Look at its study id and make a prediction out of the 4 categories. (1 prediction per image)
\n2) Look at its image id and make object detections of opacities. Here you only have 1 category -- opacity. (Could have multiple predictions per image)
There might well be some correlation between the number of opacities and what category the study is in, but you are certainly not predicting a category every time you detect an opacity.
", "votes": null, "replies": [] }, { "id": 1398704, "author_name": "hawkeat", "author_url": "", "post_date": "07/24/2021 12:15:12", "content": "My understanding is it is possible for study level to have more than one label, as each image have more than one bbox and also some study also refer to more than one images. Please correct me if I am wrong
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1397178": "Hi,\n\nI am a little confused with how the bounding boxes are labeled. While we have a study level CSV file (train_study_level.csv) which provides 4 different categories for each study, we also have image-level labeling which provides either none or opacity for bounding boxes. For images that have more than one opacity, should we assume all opacity has the same category provided by train_study_level.csv or something else? \nThe reason I am asking this is that I have noticed in https://www.kaggle.com/c/siim-covid19-detection/overview/evaluation, you can provide different labels for each opacity.", "1398222": "As far as I understand it, you're making a study level prediction for the image in the study. Then you are making a separate prediction of the opacity labels on the image. \nSo for the same image you would:\n1) Look at its study id and make a prediction out of the 4 categories. (1 prediction per image)\n2) Look at its image id and make object detections of opacities. Here you only have 1 category -- opacity. (Could have multiple predictions per image)\n\nThere might well be some correlation between the number of opacities and what category the study is in, but you are certainly not predicting a category every time you detect an opacity.", "1398704": "My understanding is it is possible for study level to have more than one label, as each image have more than one bbox and also some study also refer to more than one images. Please correct me if I am wrong" }, "source": "meta" }