{ "id": 164594, "title": "\"Wonderful\" JPG", "url": "/competitions/siim-isic-melanoma-classification/discussion/164594", "author_name": "", "post_date": "2020-07-06T19:14:37.483039600Z", "votes": 2, "comment_count": 3, "views": 0, "content": "

Maybe not constructive, but in earlier competitions, the Dicom had uncompressed data and JPG was provided as an alternative. With only JPG, it seems to be important to use various scaling of images as shown in Deotte's notebook. Or invent an affine invariant CNN... Just needed to vent.\n\"\"\n(Left: JPG, Right: f(r,g,b))

", "messages": [ { "id": "917861", "postDate": "07/06/2020 19:14:37", "content": "

Maybe not constructive, but in earlier competitions, the Dicom had uncompressed data and JPG was provided as an alternative. With only JPG, it seems to be important to use various scaling of images as shown in Deotte's notebook. Or invent an affine invariant CNN... Just needed to vent.\n\"\"\n(Left: JPG, Right: f(r,g,b))

", "rawMarkdown": "Maybe not constructive, but in earlier competitions, the Dicom had uncompressed data and JPG was provided as an alternative. With only JPG, it seems to be important to use various scaling of images as shown in Deotte's [notebook](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/160147). Or invent an affine invariant CNN... Just needed to vent.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F199197%2F4ffae9e064a50ed971655687f2ec53a1%2FKaggle.png?generation=1594062430660311&alt=media)\n(Left: JPG, Right: f(r,g,b))", "votes": null }, { "id": "918994", "postDate": "07/07/2020 16:24:19", "content": "

Hi @glimmung would you mind explaining a bit more what you are trying to show?

", "rawMarkdown": "Hi @glimmung would you mind explaining a bit more what you are trying to show?", "votes": null }, { "id": "919011", "postDate": "07/07/2020 16:39:55", "content": "

Aspolutely! I'll try. We might think the image is, e.g., 4000 x 6000, but JPG compression makes groups of 16 x 16 pixels as one \"pixel\". So, a rescaled image of 250 x 375 has information for the 16 x 16 \"pixels\", while at other parts, JPG keeps the original resolution, so you'd need the 4000 x 6000 image to harvest that information. Maybe the best analyses are good at the JPG compressed (e.g. maybe the 16 x 16 with \"belt\" are key), which would render them useless on uncompressed images?

\n\n

H'm, lots of text...

", "rawMarkdown": "Aspolutely! I'll try. We might think the image is, e.g., 4000 x 6000, but JPG compression makes groups of 16 x 16 pixels as one \"pixel\". So, a rescaled image of 250 x 375 has information for the 16 x 16 \"pixels\", while at other parts, JPG keeps the original resolution, so you'd need the 4000 x 6000 image to harvest that information. Maybe the best analyses are good at the JPG compressed (e.g. maybe the 16 x 16 with \"belt\" are key), which would render them useless on uncompressed images?\n\nH'm, lots of text...", "votes": null }, { "id": "924088", "postDate": "07/11/2020 08:48:45", "content": "

cool

", "rawMarkdown": "cool", "votes": null } ], "comments": [ { "id": 918994, "author_name": "optimo", "author_url": "", "post_date": "07/07/2020 16:24:19", "content": "

Hi @glimmung would you mind explaining a bit more what you are trying to show?

", "votes": null, "replies": [ { "id": 919011, "author_name": "glimmung", "author_url": "", "post_date": "07/07/2020 16:39:55", "content": "

Aspolutely! I'll try. We might think the image is, e.g., 4000 x 6000, but JPG compression makes groups of 16 x 16 pixels as one \"pixel\". So, a rescaled image of 250 x 375 has information for the 16 x 16 \"pixels\", while at other parts, JPG keeps the original resolution, so you'd need the 4000 x 6000 image to harvest that information. Maybe the best analyses are good at the JPG compressed (e.g. maybe the 16 x 16 with \"belt\" are key), which would render them useless on uncompressed images?

\n\n

H'm, lots of text...

", "votes": null, "replies": [] } ] }, { "id": 924088, "author_name": "", "author_url": "", "post_date": "07/11/2020 08:48:45", "content": "

cool

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "917861": "Maybe not constructive, but in earlier competitions, the Dicom had uncompressed data and JPG was provided as an alternative. With only JPG, it seems to be important to use various scaling of images as shown in Deotte's [notebook](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/160147). Or invent an affine invariant CNN... Just needed to vent.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F199197%2F4ffae9e064a50ed971655687f2ec53a1%2FKaggle.png?generation=1594062430660311&alt=media)\n(Left: JPG, Right: f(r,g,b))", "918994": "Hi @glimmung would you mind explaining a bit more what you are trying to show?", "919011": "Aspolutely! I'll try. We might think the image is, e.g., 4000 x 6000, but JPG compression makes groups of 16 x 16 pixels as one \"pixel\". So, a rescaled image of 250 x 375 has information for the 16 x 16 \"pixels\", while at other parts, JPG keeps the original resolution, so you'd need the 4000 x 6000 image to harvest that information. Maybe the best analyses are good at the JPG compressed (e.g. maybe the 16 x 16 with \"belt\" are key), which would render them useless on uncompressed images?\n\nH'm, lots of text...", "924088": "cool" }, "source": "meta" }