I want to know, is the test data set image size fixed?
\nFor data set enhancement, what should I do to prevent overfitting?
I want to know, is the test data set image size fixed?
\nFor data set enhancement, what should I do to prevent overfitting?
the test data set image size is not fixed and it is unknown, it is unknown in size and resolution.
\nIt is possible that it is kidney5 and kidney6, but not certain. Whatever code you write for the submission process should be compatible with any image size.
\nTo prevent overfitting just try to use standard data augmentations, probably just start with rotations and you could add more complex augmentations later if you see signs of overfitting
", "rawMarkdown": "the test data set image size is not fixed and it is unknown, it is unknown in size and resolution.\n\nIt is possible that it is kidney5 and kidney6, but not certain. Whatever code you write for the submission process should be compatible with any image size.\n\nTo prevent overfitting just try to use standard data augmentations, probably just start with rotations and you could add more complex augmentations later if you see signs of overfitting", "votes": null }, { "id": "2535788", "postDate": "11/23/2023 16:05:09", "content": "Thanks for your answer, but I have one more question because I'm new to the Semantic segmentation task. For this task, due to the existence of small blood vessels, I think it is inappropriate to use simple resize to change the picture to the same size and then use neural network to perform semantic segmentation. Another idea is to split the image into multiple identical pieces and feed them into the neural network to make predictions and then restore the predicted mask to the original image size. As for data enhancement, can I combine images with different resolutions into a training image similar to Mosaic data enhancement in yolov5 algorithm, since this task is to segment blood vessels anyway? Since it is a 3D kidney slice data set, should I add location coding to the training, which might have a better effect?
", "rawMarkdown": "Thanks for your answer, but I have one more question because I'm new to the Semantic segmentation task. For this task, due to the existence of small blood vessels, I think it is inappropriate to use simple resize to change the picture to the same size and then use neural network to perform semantic segmentation. Another idea is to split the image into multiple identical pieces and feed them into the neural network to make predictions and then restore the predicted mask to the original image size. As for data enhancement, can I combine images with different resolutions into a training image similar to Mosaic data enhancement in yolov5 algorithm, since this task is to segment blood vessels anyway? Since it is a 3D kidney slice data set, should I add location coding to the training, which might have a better effect?", "votes": null }, { "id": "2535823", "postDate": "11/23/2023 16:43:41", "content": "Hi Leexing, tiling the imagery for processing, then mosaicing them back together is the approach I've chosen to take for this competition. I've published the code and it's available here: https://www.kaggle.com/code/squidinator/sennet-hoa-in-memory-tiled-dataset-pytorch
", "rawMarkdown": "Hi Leexing, tiling the imagery for processing, then mosaicing them back together is the approach I've chosen to take for this competition. I've published the code and it's available here: https://www.kaggle.com/code/squidinator/sennet-hoa-in-memory-tiled-dataset-pytorch", "votes": null }, { "id": "2537108", "postDate": "11/24/2023 18:48:54", "content": "This is a very good code thank you for sharing, I was also trying to perform a tiling/mosaicing method, but I am running into the computational limit and failing submissions due to timing out. I think it is the right way to go because there is so many very small blood vessel in the image that will not be detected if i just process one 512x512 method.
\nI was thinking to maybe run a 1024x1024 single image input to account for this too, but I dont know if the computational requirements will be similar to the mosaic approach or not. Kinda just thinking out loud, but maybe you have some input on this.
", "rawMarkdown": "This is a very good code thank you for sharing, I was also trying to perform a tiling/mosaicing method, but I am running into the computational limit and failing submissions due to timing out. I think it is the right way to go because there is so many very small blood vessel in the image that will not be detected if i just process one 512x512 method.\n\nI was thinking to maybe run a 1024x1024 single image input to account for this too, but I dont know if the computational requirements will be similar to the mosaic approach or not. Kinda just thinking out loud, but maybe you have some input on this.", "votes": null } ], "comments": [ { "id": 2534732, "author_name": "peterwarren", "author_url": "", "post_date": "11/22/2023 20:40:30", "content": "the test data set image size is not fixed and it is unknown, it is unknown in size and resolution.
\nIt is possible that it is kidney5 and kidney6, but not certain. Whatever code you write for the submission process should be compatible with any image size.
\nTo prevent overfitting just try to use standard data augmentations, probably just start with rotations and you could add more complex augmentations later if you see signs of overfitting
", "votes": null, "replies": [ { "id": 2535788, "author_name": "leexing000", "author_url": "", "post_date": "11/23/2023 16:05:09", "content": "Thanks for your answer, but I have one more question because I'm new to the Semantic segmentation task. For this task, due to the existence of small blood vessels, I think it is inappropriate to use simple resize to change the picture to the same size and then use neural network to perform semantic segmentation. Another idea is to split the image into multiple identical pieces and feed them into the neural network to make predictions and then restore the predicted mask to the original image size. As for data enhancement, can I combine images with different resolutions into a training image similar to Mosaic data enhancement in yolov5 algorithm, since this task is to segment blood vessels anyway? Since it is a 3D kidney slice data set, should I add location coding to the training, which might have a better effect?
", "votes": null, "replies": [ { "id": 2535823, "author_name": "squidinator", "author_url": "", "post_date": "11/23/2023 16:43:41", "content": "Hi Leexing, tiling the imagery for processing, then mosaicing them back together is the approach I've chosen to take for this competition. I've published the code and it's available here: https://www.kaggle.com/code/squidinator/sennet-hoa-in-memory-tiled-dataset-pytorch
", "votes": null, "replies": [ { "id": 2537108, "author_name": "peterwarren", "author_url": "", "post_date": "11/24/2023 18:48:54", "content": "This is a very good code thank you for sharing, I was also trying to perform a tiling/mosaicing method, but I am running into the computational limit and failing submissions due to timing out. I think it is the right way to go because there is so many very small blood vessel in the image that will not be detected if i just process one 512x512 method.
\nI was thinking to maybe run a 1024x1024 single image input to account for this too, but I dont know if the computational requirements will be similar to the mosaic approach or not. Kinda just thinking out loud, but maybe you have some input on this.
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2534389": "I want to know, is the test data set image size fixed?\nFor data set enhancement, what should I do to prevent overfitting?", "2534732": "the test data set image size is not fixed and it is unknown, it is unknown in size and resolution.\n\nIt is possible that it is kidney5 and kidney6, but not certain. Whatever code you write for the submission process should be compatible with any image size.\n\nTo prevent overfitting just try to use standard data augmentations, probably just start with rotations and you could add more complex augmentations later if you see signs of overfitting", "2535788": "Thanks for your answer, but I have one more question because I'm new to the Semantic segmentation task. For this task, due to the existence of small blood vessels, I think it is inappropriate to use simple resize to change the picture to the same size and then use neural network to perform semantic segmentation. Another idea is to split the image into multiple identical pieces and feed them into the neural network to make predictions and then restore the predicted mask to the original image size. As for data enhancement, can I combine images with different resolutions into a training image similar to Mosaic data enhancement in yolov5 algorithm, since this task is to segment blood vessels anyway? Since it is a 3D kidney slice data set, should I add location coding to the training, which might have a better effect?", "2535823": "Hi Leexing, tiling the imagery for processing, then mosaicing them back together is the approach I've chosen to take for this competition. I've published the code and it's available here: https://www.kaggle.com/code/squidinator/sennet-hoa-in-memory-tiled-dataset-pytorch", "2537108": "This is a very good code thank you for sharing, I was also trying to perform a tiling/mosaicing method, but I am running into the computational limit and failing submissions due to timing out. I think it is the right way to go because there is so many very small blood vessel in the image that will not be detected if i just process one 512x512 method.\n\nI was thinking to maybe run a 1024x1024 single image input to account for this too, but I dont know if the computational requirements will be similar to the mosaic approach or not. Kinda just thinking out loud, but maybe you have some input on this." }, "source": "meta" }