{ "id": 334364, "title": "Question About Resizing Images", "url": "/competitions/hubmap-organ-segmentation/discussion/334364", "author_name": "", "post_date": "2022-07-01T07:20:23.768347800Z", "votes": 8, "comment_count": 3, "views": 0, "content": "

In this competition, we image Sizes (3000, 3000, 3)
\n1 - I have a question about to what extent I should reduce the size of images so that I don't lose the information for training?

\n

2- Reducing the size of the image can give a better result or a large image size can give a better result?

", "messages": [ { "id": "1839187", "postDate": "07/01/2022 07:20:23", "content": "

In this competition, we image Sizes (3000, 3000, 3)
\n1 - I have a question about to what extent I should reduce the size of images so that I don't lose the information for training?

\n

2- Reducing the size of the image can give a better result or a large image size can give a better result?

", "rawMarkdown": "In this competition, we image Sizes (3000, 3000, 3)\n1 - I have a question about to what extent I should reduce the size of images so that I don't lose the information for training?\n\n2- Reducing the size of the image can give a better result or a large image size can give a better result?", "votes": null }, { "id": "1839430", "postDate": "07/01/2022 12:06:32", "content": "

Instead of resizing you can split the image into parts, and add it to the model. just asking,

", "rawMarkdown": "Instead of resizing you can split the image into parts, and add it to the model. just asking,", "votes": null }, { "id": "1839450", "postDate": "07/01/2022 12:24:41", "content": "

Instead of resizing, you can try tiling, using tiles you can preserve high dimensional information which can be lost by resizing. please refer to this notebook.
\nIf you wish to resize, there is a method of converting the images to ZARR files, please refer to this notebook.
\nQuoting some highlights from last mentioned notebook:

\n
Super fast training: 30 minutes training...\n\n* no \"pre-tiling\" needed, the training data gets converted into .zarr files for efficient loading.\n
", "rawMarkdown": "Instead of resizing, you can try tiling, using tiles you can preserve high dimensional information which can be lost by resizing. please refer to [this](https://www.kaggle.com/code/jirkaborovec/ftus-segm-baseline-flash-unet-on-tiled-images/notebook) notebook.\nIf you wish to resize, there is a method of converting the images to ZARR files, please refer to [this](https://www.kaggle.com/code/matjes/hubmap-deepflash2-judge-price) notebook.\nQuoting some highlights from last mentioned notebook: \n```\nSuper fast training: 30 minutes training...\n\n* no \"pre-tiling\" needed, the training data gets converted into .zarr files for efficient loading.\n```", "votes": null }, { "id": "1839454", "postDate": "07/01/2022 12:28:56", "content": "

@muhammadishaque

\n

Refer the below notebook,
\nhttps://www.kaggle.com/code/muki2003/resizing-tiff-image-size

\n

You can Resize the image to any Resoluation/Size you needed. There is No loss of Information for Training !

", "rawMarkdown": "muhammadishaque \n\nRefer the below notebook,\nhttps://www.kaggle.com/code/muki2003/resizing-tiff-image-size\n\nYou can Resize the image to any Resoluation/Size you needed. There is No loss of Information for Training !", "votes": null } ], "comments": [ { "id": 1839430, "author_name": "kinnerakiran", "author_url": "", "post_date": "07/01/2022 12:06:32", "content": "

Instead of resizing you can split the image into parts, and add it to the model. just asking,

", "votes": null, "replies": [] }, { "id": 1839450, "author_name": "muhammad4hmed", "author_url": "", "post_date": "07/01/2022 12:24:41", "content": "

Instead of resizing, you can try tiling, using tiles you can preserve high dimensional information which can be lost by resizing. please refer to this notebook.
\nIf you wish to resize, there is a method of converting the images to ZARR files, please refer to this notebook.
\nQuoting some highlights from last mentioned notebook:

\n
Super fast training: 30 minutes training...\n\n* no \"pre-tiling\" needed, the training data gets converted into .zarr files for efficient loading.\n
", "votes": null, "replies": [] }, { "id": 1839454, "author_name": "muki2003", "author_url": "", "post_date": "07/01/2022 12:28:56", "content": "

@muhammadishaque

\n

Refer the below notebook,
\nhttps://www.kaggle.com/code/muki2003/resizing-tiff-image-size

\n

You can Resize the image to any Resoluation/Size you needed. There is No loss of Information for Training !

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1839187": "In this competition, we image Sizes (3000, 3000, 3)\n1 - I have a question about to what extent I should reduce the size of images so that I don't lose the information for training?\n\n2- Reducing the size of the image can give a better result or a large image size can give a better result?", "1839430": "Instead of resizing you can split the image into parts, and add it to the model. just asking,", "1839450": "Instead of resizing, you can try tiling, using tiles you can preserve high dimensional information which can be lost by resizing. please refer to [this](https://www.kaggle.com/code/jirkaborovec/ftus-segm-baseline-flash-unet-on-tiled-images/notebook) notebook.\nIf you wish to resize, there is a method of converting the images to ZARR files, please refer to [this](https://www.kaggle.com/code/matjes/hubmap-deepflash2-judge-price) notebook.\nQuoting some highlights from last mentioned notebook: \n```\nSuper fast training: 30 minutes training...\n\n* no \"pre-tiling\" needed, the training data gets converted into .zarr files for efficient loading.\n```", "1839454": "muhammadishaque \n\nRefer the below notebook,\nhttps://www.kaggle.com/code/muki2003/resizing-tiff-image-size\n\nYou can Resize the image to any Resoluation/Size you needed. There is No loss of Information for Training !" }, "source": "meta" }