I don't quite understand the last column in submission.csv, which is prediction_string. Do we have multiple masks for each image? Why, if for example we can get a mask that has all the areas that the neural network has selected. So, for example, you can do it by taking the Unet architecture. Shouldn't we build an architecture that will predict the mask separately for each found vessel?
\nI believe the first value in prediction_string is the class label. But shouldn't the desired vessel have a label of 1, not 0 as in the example? And I also understand correctly that the second element is the probability that the mask belongs to the class?
\nThank you very much in advance. I'm really confused 👀
", "messages": [ { "id": "2297181", "postDate": "06/12/2023 12:06:46", "content": "I don't quite understand the last column in submission.csv, which is prediction_string. Do we have multiple masks for each image? Why, if for example we can get a mask that has all the areas that the neural network has selected. So, for example, you can do it by taking the Unet architecture. Shouldn't we build an architecture that will predict the mask separately for each found vessel?
\nI believe the first value in prediction_string is the class label. But shouldn't the desired vessel have a label of 1, not 0 as in the example? And I also understand correctly that the second element is the probability that the mask belongs to the class?
\nThank you very much in advance. I'm really confused 👀
", "rawMarkdown": "**I don't quite understand the last column in submission.csv, which is prediction_string. Do we have multiple masks for each image? Why, if for example we can get a mask that has all the areas that the neural network has selected. So, for example, you can do it by taking the Unet architecture. Shouldn't we build an architecture that will predict the mask separately for each found vessel?**\n\n**I believe the first value in prediction_string is the class label. But shouldn't the desired vessel have a label of 1, not 0 as in the example? And I also understand correctly that the second element is the probability that the mask belongs to the class?**\n\nThank you very much in advance. I'm really confused 👀", "votes": null }, { "id": "2297245", "postDate": "06/12/2023 13:00:31", "content": "\n\nDo we have multiple masks for each image?
\n
Yes. This is an instance segmentation task.
\nIn semantic segmentation we ask the question: which pixels belong to this class?
\nIn instance segmentation we ask the question: which objects of this class exist, and which pixels belong to the instances of these objects?
\n\n\nBut shouldn't the desired vessel have a label of 1, not 0 as in the example?
\n
In this case, even though there are 3 label types provided (blood vessel, glomerules and unsure) we are only asked to predict the blood vessel class.
\n\nprobability that the mask belongs to the class
\n
The confidence that it belongs to the class, yes.
\nSee this thread for more complete answers: https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414877
It's explained on the evaluation page for this competition -
\n\"For each image in the test set, you must predict a list of instance segmentation masks and their associated detection score (Confidence). The submission.csv file uses the following format:
id,height,width,prediction_string
\nwhere prediction_string has the format 0 {confidence} {EncodedMask}. Note that the metric has several \"boilerplate\" values needed to adapt it to this competition; namely, the height, width, and the leading 0 in prediction_string, which ordinarily is a class label.\"
\nEach of the images will have a list of prediction strings that looks like this:
\n\"0 0.8376555442810059 eNqLiAgxtU2wN/M28jf0NzQwAJGoLH/DUMdA/1S7fINoIJ1iF2AY4xgAp5PtAg1jHfP9QLQpUBcaMPTHwvY38selBh/wN4SRyNjPCOpYIz9DMGnkawxkGfma+BgDcUhEjCEArAo1OQ== 0 0.7994756698608398 eNqLjMsxtkq2N/Mx8jM0MPCHkv4gbOhv6AcmgdDAAEGjsMFqgXqAwN/A0ACqHKgRBL1NfY0hLH8jX2PPiKRIYwAWVBuG 0 0.7270526885986328 eNoLis8xSTOItLd0NzQgGhj6w2gQCztGlkOWgcnBZIz8IgJCjADABx2/\"
https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/overview/evaluation
", "rawMarkdown": "It's explained on the evaluation page for this competition - \n\"For each image in the test set, you must predict a list of instance segmentation masks and their associated detection score (Confidence). The submission.csv file uses the following format: \n\nid,height,width,prediction_string \n\nwhere prediction_string has the format 0 {confidence} {EncodedMask}. Note that the metric has several \"boilerplate\" values needed to adapt it to this competition; namely, the height, width, and the leading 0 in prediction_string, which ordinarily is a class label.\"\n\nEach of the images will have a list of prediction strings that looks like this:\n\"0 0.8376555442810059 eNqLiAgxtU2wN/M28jf0NzQwAJGoLH/DUMdA/1S7fINoIJ1iF2AY4xgAp5PtAg1jHfP9QLQpUBcaMPTHwvY38selBh/wN4SRyNjPCOpYIz9DMGnkawxkGfma+BgDcUhEjCEArAo1OQ== 0 0.7994756698608398 eNqLjMsxtkq2N/Mx8jM0MPCHkv4gbOhv6AcmgdDAAEGjsMFqgXqAwN/A0ACqHKgRBL1NfY0hLH8jX2PPiKRIYwAWVBuG 0 0.7270526885986328 eNoLis8xSTOItLd0NzQgGhj6w2gQCztGlkOWgcnBZIz8IgJCjADABx2/\"\n\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/overview/evaluation", "votes": null }, { "id": "2297292", "postDate": "06/12/2023 13:41:24", "content": "Thanks so much friend! It was useful
", "rawMarkdown": "Thanks so much friend! It was useful", "votes": null }, { "id": "2297295", "postDate": "06/12/2023 13:42:23", "content": "Thanks! You helped a lot
", "rawMarkdown": "Thanks! You helped a lot", "votes": null } ], "comments": [ { "id": 2297245, "author_name": "fnands", "author_url": "", "post_date": "06/12/2023 13:00:31", "content": "\n\nDo we have multiple masks for each image?
\n
Yes. This is an instance segmentation task.
\nIn semantic segmentation we ask the question: which pixels belong to this class?
\nIn instance segmentation we ask the question: which objects of this class exist, and which pixels belong to the instances of these objects?
\n\n\nBut shouldn't the desired vessel have a label of 1, not 0 as in the example?
\n
In this case, even though there are 3 label types provided (blood vessel, glomerules and unsure) we are only asked to predict the blood vessel class.
\n\nprobability that the mask belongs to the class
\n
The confidence that it belongs to the class, yes.
\nSee this thread for more complete answers: https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414877
Thanks! You helped a lot
", "votes": null, "replies": [] } ] }, { "id": 2297249, "author_name": "saworz", "author_url": "", "post_date": "06/12/2023 13:05:32", "content": "It's explained on the evaluation page for this competition -
\n\"For each image in the test set, you must predict a list of instance segmentation masks and their associated detection score (Confidence). The submission.csv file uses the following format:
id,height,width,prediction_string
\nwhere prediction_string has the format 0 {confidence} {EncodedMask}. Note that the metric has several \"boilerplate\" values needed to adapt it to this competition; namely, the height, width, and the leading 0 in prediction_string, which ordinarily is a class label.\"
\nEach of the images will have a list of prediction strings that looks like this:
\n\"0 0.8376555442810059 eNqLiAgxtU2wN/M28jf0NzQwAJGoLH/DUMdA/1S7fINoIJ1iF2AY4xgAp5PtAg1jHfP9QLQpUBcaMPTHwvY38selBh/wN4SRyNjPCOpYIz9DMGnkawxkGfma+BgDcUhEjCEArAo1OQ== 0 0.7994756698608398 eNqLjMsxtkq2N/Mx8jM0MPCHkv4gbOhv6AcmgdDAAEGjsMFqgXqAwN/A0ACqHKgRBL1NfY0hLH8jX2PPiKRIYwAWVBuG 0 0.7270526885986328 eNoLis8xSTOItLd0NzQgGhj6w2gQCztGlkOWgcnBZIz8IgJCjADABx2/\"
https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/overview/evaluation
", "votes": null, "replies": [ { "id": 2297292, "author_name": "", "author_url": "", "post_date": "06/12/2023 13:41:24", "content": "Thanks so much friend! It was useful
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2297181": "**I don't quite understand the last column in submission.csv, which is prediction_string. Do we have multiple masks for each image? Why, if for example we can get a mask that has all the areas that the neural network has selected. So, for example, you can do it by taking the Unet architecture. Shouldn't we build an architecture that will predict the mask separately for each found vessel?**\n\n**I believe the first value in prediction_string is the class label. But shouldn't the desired vessel have a label of 1, not 0 as in the example? And I also understand correctly that the second element is the probability that the mask belongs to the class?**\n\nThank you very much in advance. I'm really confused 👀", "2297245": "> Do we have multiple masks for each image? \n\nYes. This is an instance segmentation task. \n\nIn semantic segmentation we ask the question: which pixels belong to this class? \n\nIn instance segmentation we ask the question: which objects of this class exist, and which pixels belong to the instances of these objects? \n\n> But shouldn't the desired vessel have a label of 1, not 0 as in the example? \n\nIn this case, even though there are 3 label types provided (`blood vessel`, `glomerules` and `unsure`) we are only asked to predict the `blood vessel` class. \n\n> probability that the mask belongs to the class \n\nThe confidence that it belongs to the class, yes. \nSee this thread for more complete answers: https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/414877", "2297249": "It's explained on the evaluation page for this competition - \n\"For each image in the test set, you must predict a list of instance segmentation masks and their associated detection score (Confidence). The submission.csv file uses the following format: \n\nid,height,width,prediction_string \n\nwhere prediction_string has the format 0 {confidence} {EncodedMask}. Note that the metric has several \"boilerplate\" values needed to adapt it to this competition; namely, the height, width, and the leading 0 in prediction_string, which ordinarily is a class label.\"\n\nEach of the images will have a list of prediction strings that looks like this:\n\"0 0.8376555442810059 eNqLiAgxtU2wN/M28jf0NzQwAJGoLH/DUMdA/1S7fINoIJ1iF2AY4xgAp5PtAg1jHfP9QLQpUBcaMPTHwvY38selBh/wN4SRyNjPCOpYIz9DMGnkawxkGfma+BgDcUhEjCEArAo1OQ== 0 0.7994756698608398 eNqLjMsxtkq2N/Mx8jM0MPCHkv4gbOhv6AcmgdDAAEGjsMFqgXqAwN/A0ACqHKgRBL1NfY0hLH8jX2PPiKRIYwAWVBuG 0 0.7270526885986328 eNoLis8xSTOItLd0NzQgGhj6w2gQCztGlkOWgcnBZIz8IgJCjADABx2/\"\n\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/overview/evaluation", "2297292": "Thanks so much friend! It was useful", "2297295": "Thanks! You helped a lot" }, "source": "meta" }