![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4714993%2F59b79c0f38e10c96fdc1bf52b9ab67b2%2Fheight.pnd.png?generation=1591533072911366&alt=media\")
I calculated the width and height of each image and thought it might be useful to others.
\n\nFor both width and height there seems to be some quite distinct groups. I am wondering if having different transform functions for these groups might be more beneficial. Resizing a 1000x1000 image to 768x768 isn't going to lose too much detail. However, resizing a 4000x6000 image to 768x768 may do.
\n\nMaybe something more along the lines of:\nResize to 768x768 for images originally 3000x3000 or smaller\nResize to 3000x3000 for images originally greater than 3000x3000.
\n\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4714993%2F5dfa3a173a2865e21ef22c68cdb21677%2Fwidth.png?generation=1591532976803249&alt=media\")
Notebook for those interested: https://www.kaggle.com/blueturtle/siim-list-comprehension-image-sizes/
\n\nI have attached the augmentated csv too with the height and width to save the roughly 4 hours it took the notebook to run :/
\n\nThis would require running a number of models equal to the number of splits though on a subset of the data and piecing the results together to form one final submission csv.
", "messages": [ { "id": "877235", "postDate": "06/07/2020 12:30:43", "content": "I calculated the width and height of each image and thought it might be useful to others.
\n\nFor both width and height there seems to be some quite distinct groups. I am wondering if having different transform functions for these groups might be more beneficial. Resizing a 1000x1000 image to 768x768 isn't going to lose too much detail. However, resizing a 4000x6000 image to 768x768 may do.
\n\nMaybe something more along the lines of:\nResize to 768x768 for images originally 3000x3000 or smaller\nResize to 3000x3000 for images originally greater than 3000x3000.
\n\n
Notebook for those interested: https://www.kaggle.com/blueturtle/siim-list-comprehension-image-sizes/
\n\nI have attached the augmentated csv too with the height and width to save the roughly 4 hours it took the notebook to run :/
\n\nThis would require running a number of models equal to the number of splits though on a subset of the data and piecing the results together to form one final submission csv.
", "rawMarkdown": "I calculated the width and height of each image and thought it might be useful to others.\n\nFor both width and height there seems to be some quite distinct groups. I am wondering if having different transform functions for these groups might be more beneficial. Resizing a 1000x1000 image to 768x768 isn't going to lose too much detail. However, resizing a 4000x6000 image to 768x768 may do.\n\nMaybe something more along the lines of:\nResize to 768x768 for images originally 3000x3000 or smaller\nResize to 3000x3000 for images originally greater than 3000x3000.\n\n\n\n\n\n\nNotebook for those interested: https://www.kaggle.com/blueturtle/siim-list-comprehension-image-sizes/\n\nI have attached the augmentated csv too with the height and width to save the roughly 4 hours it took the notebook to run :/ \n\nThis would require running a number of models equal to the number of splits though on a subset of the data and piecing the results together to form one final submission csv.", "votes": null }, { "id": "877299", "postDate": "06/07/2020 13:29:55", "content": "Great analysis. Thank you for this work. Processing different image sizes differently may be a good idea.
", "rawMarkdown": "Great analysis. Thank you for this work. Processing different image sizes differently may be a good idea.", "votes": null }, { "id": "877633", "postDate": "06/07/2020 19:30:20", "content": "great idea, would love to give it a try.
", "rawMarkdown": "great idea, would love to give it a try.", "votes": null }, { "id": "882464", "postDate": "06/11/2020 20:19:00", "content": "Nice. I guess this is how the leak is searched ;)
", "rawMarkdown": "Nice. I guess this is how the leak is searched ;)", "votes": null } ], "comments": [ { "id": 877299, "author_name": "cdeotte", "author_url": "", "post_date": "06/07/2020 13:29:55", "content": "Great analysis. Thank you for this work. Processing different image sizes differently may be a good idea.
", "votes": null, "replies": [] }, { "id": 877633, "author_name": "rohitsingh9990", "author_url": "", "post_date": "06/07/2020 19:30:20", "content": "great idea, would love to give it a try.
", "votes": null, "replies": [] }, { "id": 882464, "author_name": "mks2192", "author_url": "", "post_date": "06/11/2020 20:19:00", "content": "Nice. I guess this is how the leak is searched ;)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "877235": "I calculated the width and height of each image and thought it might be useful to others.\n\nFor both width and height there seems to be some quite distinct groups. I am wondering if having different transform functions for these groups might be more beneficial. Resizing a 1000x1000 image to 768x768 isn't going to lose too much detail. However, resizing a 4000x6000 image to 768x768 may do.\n\nMaybe something more along the lines of:\nResize to 768x768 for images originally 3000x3000 or smaller\nResize to 3000x3000 for images originally greater than 3000x3000.\n\n\n\n\n\n\nNotebook for those interested: https://www.kaggle.com/blueturtle/siim-list-comprehension-image-sizes/\n\nI have attached the augmentated csv too with the height and width to save the roughly 4 hours it took the notebook to run :/ \n\nThis would require running a number of models equal to the number of splits though on a subset of the data and piecing the results together to form one final submission csv.", "877299": "Great analysis. Thank you for this work. Processing different image sizes differently may be a good idea.", "877633": "great idea, would love to give it a try.", "882464": "Nice. I guess this is how the leak is searched ;)" }, "source": "meta" }