{ "id": 333552, "title": " is this a mistake ? 6.263 µm for prostate", "url": "/competitions/hubmap-organ-segmentation/discussion/333552", "author_name": "", "post_date": "2022-06-27T07:34:48.216936900Z", "votes": 13, "comment_count": 12, "views": 0, "content": "

It is mentioned in the data page that
\nHPA : 0.400 µm for prostate
\nHubmap: 6.263 µm for prostate

\n

6.263/0.400 = 15
\nif you scale 15x smaller for 3000x3000 train image, you end up 200x200.
\nit looks very small. is this correct?

", "messages": [ { "id": "1834793", "postDate": "06/27/2022 07:34:48", "content": "

It is mentioned in the data page that
\nHPA : 0.400 µm for prostate
\nHubmap: 6.263 µm for prostate

\n

6.263/0.400 = 15
\nif you scale 15x smaller for 3000x3000 train image, you end up 200x200.
\nit looks very small. is this correct?

", "rawMarkdown": "It is mentioned in the data page that\nHPA : 0.400 µm for prostate\nHubmap: 6.263 µm for prostate\n\n6.263/0.400 = 15\nif you scale 15x smaller for 3000x3000 train image, you end up 200x200.\nit looks very small. is this correct?", "votes": null }, { "id": "1835300", "postDate": "06/27/2022 16:31:07", "content": "

From the data description: The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.

", "rawMarkdown": "From the data description: `The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.`", "votes": null }, { "id": "1835446", "postDate": "06/27/2022 18:22:50", "content": "

Just to confirm, as your answer isn't 100% clear.

\n

This is not a typo? One pixel in prostate scans occupies a real-world space (2D) of 6.263x6.263µm?

", "rawMarkdown": "Just to confirm, as your answer isn't 100% clear. \n\nThis is not a typo? One pixel in prostate scans occupies a real-world space (2D) of 6.263x6.263µm?", "votes": null }, { "id": "1835459", "postDate": "06/27/2022 18:30:48", "content": "

\"https://i.ibb.co/hRQ7t9d/Selection-064.png\"

\n

i think 6.2630 um per pixel resolution make image too small. Hence I suspect it is a typo error (or 6.x seems to be sensor size).

\n

usually we are looking at 20x magification

\n

\"https://i.ibb.co/p3KmQSg/Selection-065.png\"

", "rawMarkdown": "![https://i.ibb.co/hRQ7t9d/Selection-064.png](https://i.ibb.co/hRQ7t9d/Selection-064.png)\n\ni think 6.2630 um per pixel resolution make image too small. Hence I suspect it is a typo error (or 6.x seems to be sensor size).\n\nusually we are looking at 20x magification\n\n![https://i.ibb.co/p3KmQSg/Selection-065.png](https://i.ibb.co/p3KmQSg/Selection-065.png)", "votes": null }, { "id": "1835513", "postDate": "06/27/2022 19:34:37", "content": "

@yashvrdnjain – please comment on this

", "rawMarkdown": "yashvrdnjain – please comment on this", "votes": null }, { "id": "1835517", "postDate": "06/27/2022 19:44:33", "content": "

Here's another photo showing the difference:

\n

3000x3000-->191x191 (0.4-->6.263)

\n

\"comparison_img\"

", "rawMarkdown": "Here's another photo showing the difference:\n\n3000x3000-->191x191 (0.4-->6.263)\n\n![comparison_img](https://i.ibb.co/R4cK7V6/comparison-1.png)", "votes": null }, { "id": "1835550", "postDate": "06/27/2022 20:36:45", "content": "

I've verified with the hosts that there is no typo the resolution of Hubmap prostate images is 6.263 um.

", "rawMarkdown": "I've verified with the hosts that there is no typo the resolution of Hubmap prostate images is 6.263 um.", "votes": null }, { "id": "1835556", "postDate": "06/27/2022 20:49:08", "content": "

Ok. Thank you for clarifying.

", "rawMarkdown": "Ok. Thank you for clarifying.", "votes": null }, { "id": "1837473", "postDate": "06/29/2022 16:12:18", "content": "

6.263 um is probably correct (though i find it strange)
\nother kaggler may want to verify and report results.
\nthis is what i have:

\n
inference code: LB 0.46\n\nd = df.iloc[index]\npixel_size = d.pixel_size\nif d.organ=='prostate':\n        pixel_size = 0.4\n\ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n
\n
inference code: LB 0.56\n\nd = df.iloc[index]\npixel_size = d.pixel_size \n\ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n
", "rawMarkdown": "6.263 um is probably correct (though i find it strange)\nother kaggler may want to verify and report results.\nthis is what i have:\n\n```\ninference code: LB 0.46\n\nd = df.iloc[index]\npixel_size = d.pixel_size\nif d.organ=='prostate':\n pixel_size = 0.4\n \ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n \n```\n\n```\ninference code: LB 0.56\n\nd = df.iloc[index]\npixel_size = d.pixel_size \n \ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n \n```", "votes": null }, { "id": "1851791", "postDate": "07/11/2022 14:41:31", "content": "

It's probably not a typo but they did it intentionally. In one of organizer's posts, he said that generalization is the most important goal of this competition, so they might not used magnification on purpose. However providing only one scale per source for every organ didn't make sense to me. I would prepare images on multiple scales.

", "rawMarkdown": "It's probably not a typo but they did it intentionally. In one of organizer's posts, he said that generalization is the most important goal of this competition, so they might not used magnification on purpose. However providing only one scale per source for every organ didn't make sense to me. I would prepare images on multiple scales.", "votes": null }, { "id": "1883878", "postDate": "08/04/2022 06:47:11", "content": "

One question,

\n
\n

s = pixel_size/0.4 * TRAINING_SCALE
\n image = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=…)

\n
\n

This step attempts to normalize the test tiff images to pixel_size = 0.4um level right, then we can perform window slice or directly resize the output image. Correct?

", "rawMarkdown": "One question, \n\n> s = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n\nThis step attempts to normalize the test tiff images to `pixel_size = 0.4um` level right, then we can perform window slice or directly resize the output `image`. Correct?", "votes": null }, { "id": "1883922", "postDate": "08/04/2022 07:35:52", "content": "

yes. that is right

", "rawMarkdown": "yes. that is right", "votes": null }, { "id": "1883948", "postDate": "08/04/2022 08:01:04", "content": "

Thanks @hengck23 , you're always so helpful!😁

", "rawMarkdown": "Thanks @hengck23 , you're always so helpful!😁", "votes": null } ], "comments": [ { "id": 1835300, "author_name": "sohier", "author_url": "", "post_date": "06/27/2022 16:31:07", "content": "

From the data description: The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.

", "votes": null, "replies": [ { "id": 1835446, "author_name": "dschettler8845", "author_url": "", "post_date": "06/27/2022 18:22:50", "content": "

Just to confirm, as your answer isn't 100% clear.

\n

This is not a typo? One pixel in prostate scans occupies a real-world space (2D) of 6.263x6.263µm?

", "votes": null, "replies": [] }, { "id": 1835459, "author_name": "hengck23", "author_url": "", "post_date": "06/27/2022 18:30:48", "content": "

\"https://i.ibb.co/hRQ7t9d/Selection-064.png\"

\n

i think 6.2630 um per pixel resolution make image too small. Hence I suspect it is a typo error (or 6.x seems to be sensor size).

\n

usually we are looking at 20x magification

\n

\"https://i.ibb.co/p3KmQSg/Selection-065.png\"

", "votes": null, "replies": [] }, { "id": 1835513, "author_name": "dschettler8845", "author_url": "", "post_date": "06/27/2022 19:34:37", "content": "

@yashvrdnjain – please comment on this

", "votes": null, "replies": [] }, { "id": 1835517, "author_name": "dschettler8845", "author_url": "", "post_date": "06/27/2022 19:44:33", "content": "

Here's another photo showing the difference:

\n

3000x3000-->191x191 (0.4-->6.263)

\n

\"comparison_img\"

", "votes": null, "replies": [] }, { "id": 1835550, "author_name": "sohier", "author_url": "", "post_date": "06/27/2022 20:36:45", "content": "

I've verified with the hosts that there is no typo the resolution of Hubmap prostate images is 6.263 um.

", "votes": null, "replies": [] }, { "id": 1835556, "author_name": "dschettler8845", "author_url": "", "post_date": "06/27/2022 20:49:08", "content": "

Ok. Thank you for clarifying.

", "votes": null, "replies": [] }, { "id": 1837473, "author_name": "hengck23", "author_url": "", "post_date": "06/29/2022 16:12:18", "content": "

6.263 um is probably correct (though i find it strange)
\nother kaggler may want to verify and report results.
\nthis is what i have:

\n
inference code: LB 0.46\n\nd = df.iloc[index]\npixel_size = d.pixel_size\nif d.organ=='prostate':\n        pixel_size = 0.4\n\ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n
\n
inference code: LB 0.56\n\nd = df.iloc[index]\npixel_size = d.pixel_size \n\ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n
", "votes": null, "replies": [] }, { "id": 1883878, "author_name": "fuckvenkatraman", "author_url": "", "post_date": "08/04/2022 06:47:11", "content": "

One question,

\n
\n

s = pixel_size/0.4 * TRAINING_SCALE
\n image = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=…)

\n
\n

This step attempts to normalize the test tiff images to pixel_size = 0.4um level right, then we can perform window slice or directly resize the output image. Correct?

", "votes": null, "replies": [] }, { "id": 1883922, "author_name": "hengck23", "author_url": "", "post_date": "08/04/2022 07:35:52", "content": "

yes. that is right

", "votes": null, "replies": [] }, { "id": 1883948, "author_name": "fuckvenkatraman", "author_url": "", "post_date": "08/04/2022 08:01:04", "content": "

Thanks @hengck23 , you're always so helpful!😁

", "votes": null, "replies": [] } ] }, { "id": 1851791, "author_name": "gunesevitan", "author_url": "", "post_date": "07/11/2022 14:41:31", "content": "

It's probably not a typo but they did it intentionally. In one of organizer's posts, he said that generalization is the most important goal of this competition, so they might not used magnification on purpose. However providing only one scale per source for every organ didn't make sense to me. I would prepare images on multiple scales.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1834793": "It is mentioned in the data page that\nHPA : 0.400 µm for prostate\nHubmap: 6.263 µm for prostate\n\n6.263/0.400 = 15\nif you scale 15x smaller for 3000x3000 train image, you end up 200x200.\nit looks very small. is this correct?", "1835300": "From the data description: `The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.`", "1835446": "Just to confirm, as your answer isn't 100% clear. \n\nThis is not a typo? One pixel in prostate scans occupies a real-world space (2D) of 6.263x6.263µm?", "1835459": "![https://i.ibb.co/hRQ7t9d/Selection-064.png](https://i.ibb.co/hRQ7t9d/Selection-064.png)\n\ni think 6.2630 um per pixel resolution make image too small. Hence I suspect it is a typo error (or 6.x seems to be sensor size).\n\nusually we are looking at 20x magification\n\n![https://i.ibb.co/p3KmQSg/Selection-065.png](https://i.ibb.co/p3KmQSg/Selection-065.png)", "1835513": "yashvrdnjain – please comment on this", "1835517": "Here's another photo showing the difference:\n\n3000x3000-->191x191 (0.4-->6.263)\n\n![comparison_img](https://i.ibb.co/R4cK7V6/comparison-1.png)", "1835550": "I've verified with the hosts that there is no typo the resolution of Hubmap prostate images is 6.263 um.", "1835556": "Ok. Thank you for clarifying.", "1837473": "6.263 um is probably correct (though i find it strange)\nother kaggler may want to verify and report results.\nthis is what i have:\n\n```\ninference code: LB 0.46\n\nd = df.iloc[index]\npixel_size = d.pixel_size\nif d.organ=='prostate':\n pixel_size = 0.4\n \ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n \n```\n\n```\ninference code: LB 0.56\n\nd = df.iloc[index]\npixel_size = d.pixel_size \n \ns = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n \n```", "1851791": "It's probably not a typo but they did it intentionally. In one of organizer's posts, he said that generalization is the most important goal of this competition, so they might not used magnification on purpose. However providing only one scale per source for every organ didn't make sense to me. I would prepare images on multiple scales.", "1883878": "One question, \n\n> s = pixel_size/0.4 * TRAINING_SCALE\nimage = cv2.resize(image,dsize=None,fx=s,fy=s,interpolation=...)\n\nThis step attempts to normalize the test tiff images to `pixel_size = 0.4um` level right, then we can perform window slice or directly resize the output `image`. Correct?", "1883922": "yes. that is right", "1883948": "Thanks @hengck23 , you're always so helpful!😁" }, "source": "meta" }